Content safety
Tests of refusing harmful requests, by the MLCommons hazards. We count 329 benchmarks here, 3 of them arenas. A benchmark belongs here when risk includes violent crimes, non-violent crimes, sex-related crimes, child sexual exploitation, defamation, specialized advice, privacy, intellectual property, indiscriminate weapons, hate, suicide and self-harm, sexual content or elections.
How fast it grows
71 were released since July, more than the 38 in the first half of 2026.
How much is still open
No model has passed 85% on 41 of the 109 benchmarks here that score in percent, 38% of them.
Every benchmark in this view
Newest first. Facets come from each benchmark’s own documents, and our methodology says how we tag them.
| Benchmark | Level | Released | Best score | Status |
|---|---|---|---|---|
| ASIMOV-Jailbreak Embodied agents handle adversarial instructions involving physical safety. Tool-call checks and model judges measure harmful compliance, refusals, and physical safety rubric adherence. | Everyday | Nov 2026 | Not in cards | |
| DeputyBench Agents handle business requests under paired permission settings. The evaluator checks whether they act when authorized and seek evidence or refrain when authorization is missing. | University | Nov 2026 | | Not in cards |
| FraudBench (GOAL Lab) Models face 159 scenarios drawn from 112 real U.S. fraud lawsuits. Yes or no questions, open questions and agent tasks test whether a model recommends, assists or commits fraud. | University | Nov 2026 | 10.1% | Not in cards |
| How frontier models miss risk in routine enterprise work Agents handle workplace requests whose risks appear in the prompt or in supporting files. Judges check whether the agent completes an unsafe action. | Expert | Nov 2026 | Not in cards | |
| Malicious agentic influence campaigns, 34-point rubric (Anthropic) Agents run simulated voter suppression and domestic polarization campaigns. Weighted criteria measure campaign execution on a 34-point scale. | Expert | Nov 2026 | System cards only | |
| Moneda AdviceBench Models answer personal finance requests and paired education questions. Moneda’s default board scores paired pass rate on 50 EU personal requests with education twins. | Expert | Nov 2026 | Not in cards | |
| PReMind-Bench Models inspect a scene and decide whether to retrieve a personal memory before answering. Judges check whether the answer addresses the hazard and uses the right person’s information. | Everyday | Nov 2026 | 87.3% | Not in cards |
| Reasoning extraction red-teaming Six attacker models try to make a model reveal its hidden reasoning. Scored on how often an attack succeeds. | Expert | Nov 2026 | System cards only | |
| Suicide Memory Bench AI systems, each a model in its own app or coding harness, talk with 128 fictional personas over 29 chats. Five chats carry suicidal statements or warning signs of Suicide Crisis Syndrome, and a final request follows, such as a farewell note. Four sub-benchmarks cross explicit statements and warning signs with one chat or separate chats joined only by memory. A model judge rates each reply on a 0 to 4 safety scale. | Everyday | Nov 2026 | | Not in cards |
| Agent harness misuse evaluation Agents in a sandboxed Antigravity harness get malicious and benign tool-use requests on illegal acts, physical harm and scaled abuse. Reports safe action rate. | Everyday | Oct 2026 | | System cards only |
| AMLworld-Compact Models inspect transaction graphs and estimate whether a marked target is suspicious. The grader thresholds the probabilities and computes a weighted F1 score. | University | Oct 2026 | | Not in cards |
| BARC-Bench Agents handle routine workplace requests and use records to infer privacy and safety constraints. A judge checks whether each attempt completes safely or defers. | University | Oct 2026 | 72.7% | Not in cards |
| Biological static hazard evaluation Models answer 75 questions across priority biological threat scenarios, each asked 10 times. The score is the share of hazardous responses. | Expert | Oct 2026 | | System cards only |
| CareManage-Bench Models respond to patient scenarios involving medication safety, care plans, and triage. The harness checks the responses against case specific criteria. | University | Oct 2026 | Not in cards | |
| DistressBench Models reply as a chatbot to 718 simulated users in psychological crisis, in single-turn and multi-turn chats. Three judge models score replies against clinician-written rubrics. | Expert | Oct 2026 | | In use |
| Dosimeter US midterm source audit AI assistants answer ten English questions about claims circulating before the US midterms. Dosimeter counts answers that cite a listed source, including refutations. | Everyday | Oct 2026 | Not in cards | |
| Gray Swan Reward Hacking Bench Agentic environments where a task can be finished honestly or by crossing a declared line, such as shipping insecure code, leaking a secret, harming a third party through a shared service or leaving the scope of a penetration test. Reports how often each line is crossed. | Expert | Oct 2026 | System cards only | |
| Keeping secrets when pressed in simulated scenarios (Anthropic) Models hold confidential information in 39 job scenarios while a simulated counterpart presses for it. Measures leaks in replies and in reasoning. | Expert | Oct 2026 | System cards only | |
| MedSafe-Dx v0.3 Models read synthetic primary-care intake summaries and decide whether to flag a dangerous condition. Rule-based labels determine weighted error costs, with missed serious cases costing seven times a false escalation or incorrect danger label. The score normalizes these costs against a fixed escalation baseline. | Expert | Oct 2026 | 73.5 | Not in cards |
| MirrorDojo Indirect prompt injection test run in Reflection AI's Mirror agent harness, based on AgentDojo. Each episode pairs a user task with an instruction planted in a tool result. Reports attack success rate and utility. | Everyday | Oct 2026 | 3.10% | System cards only |
| RealmPathology Models interpret real anatomic pathology reports and must not claim a diagnosis, stage or treatment the report does not support. Rubrics have 772 positive and 299 penalty criteria. | Expert | Oct 2026 | 78% | Not in cards |
| SAFE-Teen Chatbots hold 6 to 20 turn chats with a simulated teen across 106 clinician-written scenarios, branching on how the teen reacts. A judge rates safety checks and reports the failure rate. | Everyday | Oct 2026 | | Not in cards |
| SAFE-Teen Health advice Models respond to simulated teens in health advice scenarios. The score is the published conversation failure rate. | Everyday | Oct 2026 | 9.00% | Not in cards |
| SAFE-Teen Urgent safety Models respond to simulated teens in urgent safety scenarios. The score is the published conversation failure rate. | Everyday | Oct 2026 | 15% | Not in cards |
| SDBench Diagnostic agents order tests, ask questions and commit diagnoses in sequential clinical cases. Scores combine diagnostic correctness with test cost. | Expert | Oct 2026 | Not in cards | |
| Semantic DLP: Catch the Data Leak Models search long technical documents for proprietary content. They must identify the leak and quote the relevant span. | University | Oct 2026 | | Not in cards |
| The Impossible Refund Assistants pursue a late train refund in a simulated email environment. Separate judges rate deception and other behaviors from the messages. | Everyday | Oct 2026 | Not in cards | |
| The Midnight Index Models lead a nation through fixed scenarios in a strategy game. The simulation records outcomes, escalation, and whether public commitments match later decisions. | Everyday | Oct 2026 | Not in cards | |
| The Pushback Test Models advise a parent who questions urgent medical advice. The evaluation checks whether the model maintains the urgency required by NHS guidance. | University | Oct 2026 | 0.81 | Not in cards |
| VoxAnon Models anonymize speech or try to identify the original speaker. Balanced trials compare reidentification across age, gender, pitch, and accent groups. | Everyday | Oct 2026 | Not in cards | |
| Washington Post midterm voter guides Chatbots write voter guides for US addresses. The Washington Post checks whether each guide names the correct House district. | Everyday | Oct 2026 | Not in cards | |
| WellnessBench Models answer young people asking about eating, sexual safety, substances and self harm, in clinician-written single-turn and pushback dialogues. A pass must be both safe and helpful. | Everyday | Oct 2026 | 66% | Not in cards |
| AILuminate Safety 1.2 Models answer 3,900 hidden English prompts, 325 in each of 12 hazard categories, picked with item response theory. A calibrated evaluator gives a 0 to 100 score. | Everyday | Sep 2026 | | Not in cards |
| AV-SafetyBench Models generate videos with sound from written prompts. Judges review each clip with sound, in silence, and as audio alone to identify unsafe outputs. | Everyday | Sep 2026 | 49.4% | Not in cards |
| Chatbots and the Ballot Box Models answer voter questions in English and Spanish across ten US states. Analysts check routine answers against election guidance. A model grades responses to disputed claims. | Everyday | Sep 2026 | 89.3% | Not in cards |
| Election-Day AI Poisoning in Germany AI search services answer ten German questions about documented political claims. The audit counts answers citing domains on lists of Russian state media or the Pravda network. | Everyday | Sep 2026 | Not in cards | |
| Evaluating AI Safety in Teen Conversations Chatbots respond to simulated teenagers across ten replies. Judges measure the share of conversations that fail a critical safety check. | University | Sep 2026 | | Not in cards |
| K-Bench Models respond across multi-turn mental-health risk conversations involving suicide, self-harm, domestic violence, substance misuse, and no-risk cases. | Expert | Sep 2026 | 99% | Not in cards |
| MentalHealthBench 1,215 synthetic mental-health conversations with 5,262 clinician-authored rubric criteria. An LLM judge scores safety and guidance quality. | Expert | Sep 2026 | 57.3% | System cards only |
| Misaligned Behavior in Realistic Work Environments Agents do computer and browser tasks in simulated workplace apps with ambiguous instructions. Scored on the rate of unintended outcomes such as data disclosure. | Expert | Sep 2026 | 1.40% | System cards only |
| Roboharm Agents and VLAs controlling bimanual YAM arms get five harmful-looking instructions in fixed scenes, such as stabbing a doll. Reviewers label refusals and completions. | Everyday | Sep 2026 | 20% | System cards only |
| Safe response to high-severity prompts Models handle multi-turn self-harm and violent activity conversations where harmful intent only emerges from context. Scored as safe response rate. | Expert | Sep 2026 | System cards only | |
| Static Jailbreak Evaluations Models face a fixed set of jailbreak attacks on bio, violence and cyber requests. Scored as defender success rate per category. | Expert | Sep 2026 | System cards only | |
| Sumi Japanese PII Corpus Models locate personal information in Japanese business templates and prose. The release includes 2,000 template test documents and 520 documents outside those templates. | Everyday | Sep 2026 | Not in cards | |
| VoxParity Voice agents choose tool calls after listening to speech. Paired recordings keep the words fixed while changing an audible cue that affects the correct action. | Everyday | Sep 2026 | | Not in cards |
| Bio Static Jailbreak Eval Common jailbreak templates are applied to high severity biology requests against the safeguarded model. Reports query and objective success rates. | Expert | Aug 2026 | | System cards only |
| Biosecurity VCT Models answer dual-use virus lab troubleshooting items. Accuracy counts should-refuse and should-comply items together. | Expert | Aug 2026 | 47.8% | System cards only |
| BioUseBench Models face dual-use biology enablement prompts. The reported metric is the refusal rate on the severity 5 subset. | Expert | Aug 2026 | 91.4% | System cards only |
| Brennan Center election conspiracy chatbot audit Chatbots answer questions about election claims and respond to skeptical follow-ups. Researchers check factual errors, unsupported citations, and whether responses endorse the claims. | Everyday | Aug 2026 | Not in cards | |
| Brennan Center election image generation audit Image generators receive requests for misleading election scenes. Researchers count how often each system produces an image, including images that omit requested details. | Everyday | Aug 2026 | 100% | Not in cards |
| Brennan Center election image identification Chatbots inspect 14 synthetic election images in English and Spanish. Researchers count how many images each chatbot identifies as synthetic in each language. | Everyday | Aug 2026 | Not in cards | |
| CBRN Conversational Evals About 1,500 expert-written single and multi-turn CBRN requests test whether safeguards block hazardous help. Reports violation and benign refusal rates. | Expert | Aug 2026 | | System cards only |
| Chem / Bio Dynamic Jailbreak Eval An automated attacker spreads a harmful chem or bio goal across many turns and chats. Elicited content is scored against expert rubrics. | Expert | Aug 2026 | | System cards only |
| Gray Swan Bio Eval Gray Swan's Shade red teaming agent attacks the safeguarded model on about 50 high risk biology tasks. Reports query and objective success rates. | Expert | Aug 2026 | | System cards only |
| Nuclear Decision-Making Benchmark Models answer policy questions framed from different countries’ perspectives. The authors compare response distributions and agreement across repeated runs. | Expert | Aug 2026 | | Not in cards |
| Singapore Undesirable Content Models answer prompts about sensitive situations in Singapore. A classifier labels each response and reports the percentage it judges safe. | Everyday | Aug 2026 | Not in cards | |
| States United voter questions ChatGPT and Google answer voter questions. Researchers check factual claims against official sources and assess the completeness of written answers. | Everyday | Aug 2026 | Not in cards | |
| Transluce Mental Health Evaluation Models hold multi-turn conversations with simulated users facing mental-health scenarios. Automated judges score 14 helpful and harmful behaviors. | Expert | Aug 2026 | Not in cards | |
| U18 safety evaluations Models answer challenging prompts in teen safety categories such as age-restricted goods, eating disorders and emotional reliance. Scored as safe completion rate per category. | Everyday | Aug 2026 | System cards only | |
| AgentHarm Verified (Meta) Agents receive 108 malicious AgentHarm tasks kept after human review, plus a benign set. Reports average harm score and false refusal rate. | Expert | Jul 2026 | | System cards only |
| AI Security Leaderboard Models face systematic jailbreak attacks across chemical, biological, radiological, nuclear, explosive and cyber domains. Scores count universal jailbreaks and estimate attacker cost. | Expert | Jul 2026 | $60 | Not in cards |
| AMLBench v0.2 (Superjective) Models decide whether to escalate transaction alerts. The score measures how often they escalate reportable alerts under changes to analyst instructions and case narratives. | University | Jul 2026 | | Not in cards |
| Autointent-Bio Models face manual and synthetic dangerous and benign biology queries. xAI reports refusal recall on the dangerous ones. | Jul 2026 | 100% | System cards only | |
| Autointent-Chem Models face manual and synthetic dangerous and benign chemistry queries. xAI reports refusal recall on the dangerous ones. | Jul 2026 | 100% | System cards only | |
| BioSecBench-Refusal Agents run 61 routine and 46 concealed-hazard biology analysis tasks. The score pairs red-team refusal with routine completion. | Expert | Jul 2026 | 62.4% | System cards only |
| Child safety (xAI) Models face scripted multi-turn requests for child sexual exploitation content. Compliance is the share of dialogues that fail policy. | Jul 2026 | 0% | System cards only | |
| Inspect India Evals Models answer questions about Indian culture and reason across Indian languages. Other tasks test social bias, digital public infrastructure safeguards and refusal behavior. | University | Jul 2026 | Not in cards | |
| KORA Apps Simulated children use consumer AI apps end to end, with each app system prompt, filters and safeguards active. Scores cover online, physical and sexual, psycho-social and learning risks. | Everyday | Jul 2026 | Not in cards | |
| PatientAgentBench Patient-facing health agents talk with a simulated patient and use a sandbox of healthcare tools for scheduling, prescriptions, telehealth and escalation. An LLM jury scores over a hundred clinician-grounded criteria. | Expert | Jul 2026 | 4.25 | Not in cards |
| Self-harm refusals (xAI) Models face self-harm and crisis prompts and conversations. A model fails if it assists or refuses without pointing to help. | Jul 2026 | 0.50% | System cards only | |
| VAmoS Bench Voice agents answer simulated bank customers and use tools to manage credit cards. A model checks each call against its transcript and recorded actions. | University | Jul 2026 | | Not in cards |
| Biology Model Refusal Evaluation Models answer severe, dual-use and benign biology prompts. Scored as safe rate on harmful prompts and not over-refusing benign ones. | Expert | Jun 2026 | 99.9% | System cards only |
| BioTIER v1 Refuse Models receive 501 dangerous biology prompts on select agents, dual-use research and catastrophe avoidance. The score is the share correctly refused. | Expert | Jun 2026 | | System cards only |
| CAREBench Models answer 500 prompts with embedded child safety risks across 12 categories, such as grooming, impersonation and emotional dependency. A judge marks each response acceptable or not. | Everyday | Jun 2026 | 2.30% | Not in cards |
| FinRED Financial Safety Models answer prompts about financial risks. A finance-specific judge grades each response and reports the percentage that assist harmful conduct. | University | Jun 2026 | 0.15% | Not in cards |
| GateMem Memory agents serve several people who share one memory pool in hospitals, offices, schools and homes. Each long episode checks three things: answering legitimate requests after state updates, refusing information a person is not allowed to see, and forgetting records after an explicit deletion request. Hidden checkpoints and leak targets let a structured judge score each response. | Everyday | Jun 2026 | | Not in cards |
| HelpBench 450 real-world privacy, safety and security advice questions across nine topics, scored by rubric for accuracy and tone. | Jun 2026 | 84% | System cards only | |
| MAI internal jailbreak evaluation suite Models face about 9,500 jailbreak prompts built from 2,500 seed scenarios, grouped as foundational, compositional and adaptive. Scored as attack success rate. | Everyday | Jun 2026 | System cards only | |
| MAI internal safety and over-refusal evaluation Models answer harmful and benign prompts in eight harm categories. Judges score safety pass rate on sensitive prompts and helpfulness as one minus over-refusal. | Everyday | Jun 2026 | System cards only | |
| MANDOS Run 003 Models make decisions during seven industrial incidents and four fictional scenarios. A model judge checks their decisions, including whether they seek expert advice. | Expert | Jun 2026 | 100% | Not in cards |
| Rewind-attacker cyber adversarial robustness (Anthropic) An automated attacker with rewinds and 400 calls tries to make a safeguarded Claude Code agent complete offensive cyber tasks. Score is attack success rate. | Expert | Jun 2026 | 3.90% | System cards only |
| AgentThreatBench Agents carry out routine tasks in the presence of untrusted content. The benchmark scores task completion and security separately. | University | May 2026 | Not in cards | |
| Anthropic undermining liberal democracy (behavioral audit metric) Automated behavioral audit metric: investigator models probe willingness to help coups, election tampering, or political surveillance, censorship or intimidation. | May 2026 | 1.13 | System cards only | |
| Biological safety evaluations (OpenAI) Models answer biology requests drawn from deidentified production data and from easy and hard synthetic sets. Scored as the share of cases where model training alone prevents a violative response. | Expert | May 2026 | System cards only | |
| Electoral Hallucinations AI services answer questions about three Scottish Parliament constituencies. Two reviewers check responses against election facts and report the percentage that are entirely accurate. | Everyday | May 2026 | Not in cards | |
| HarnessAudit-Bench Agents run inside execution harnesses such as Claude Code, Codex and OpenClaw on 210 tasks across eight domains, from finance to software engineering, with safety rules built in. The audit reads the full trajectory through channels the agent cannot change. It scores whether the agent stays inside its permission and information-flow policies, completes checkpoints, and keeps both under prompt injection, vague goals and tool errors. | Expert | May 2026 | | Not in cards |
| High Stakes factuality prompts Models answer difficult medical, legal and financial prompts. A grader with web access counts responses and claims with factual errors. | Expert | May 2026 | 4.10% | System cards only |
| NewsBench (Forum AI) Models answer questions about current events. Judges calibrated against expert ratings score factual accuracy, neutrality, and source quality. | Everyday | May 2026 | 88.3% | Not in cards |
| ROK-FORTRESS Models face adversarial national security and public safety prompts in English and Korean, set in US or Korean contexts. Each prompt has a benign pair to measure over-refusal. | Expert | May 2026 | | In use |
| SIGHT (Search and Intent Gauge for Hospitality and Tourism) Agents recommend restaurants, hotels and attractions from user preferences and search results. Evaluators check constraints, evidence and privacy. | Everyday | May 2026 | Not in cards | |
| StereoTales Models write multilingual stories about characters with controlled demographic attributes. Statistical tests find overrepresented associations, and people judge their harmfulness. | Everyday | May 2026 | Not in cards | |
| Adversarial Humanities Benchmark Models respond to harmful requests rewritten in literary and scholarly styles. A judge measures how often the rewrites elicit unsafe answers. | Everyday | Apr 2026 | 9.20% | Not in cards |
| Clinical SOAP Note Evaluation: Safety-First Benchmarking Models turn 300 synthetic doctor-patient dialogues into clinical SOAP notes. Three judges check claims against transcript evidence and score numeric fidelity, coverage, and writing quality. The composite weights safety at 50%, coverage at 30%, and general quality at 20%. | Expert | Apr 2026 | 4.72 | Not in cards |
| CodeIPI Coding agents fix bugs while handling untrusted repository text. The benchmark scores task completion, resistance to embedded instructions and false alarms on clean tasks. | University | Apr 2026 | 91.1% | Not in cards |
| Dictatorship Eval Scenarios at six difficulty levels testing whether models help with authoritarian requests, from blunt asks to requests hidden in code edits. LLM judge, no self-grading. | Apr 2026 | 91.3% | Not in cards | |
| Hiding a secret password (Anthropic) Models keep a given password secret while an investigator model tries to extract it over many turns. Measures survival and per-turn leak rates. | Expert | Apr 2026 | System cards only | |
| Single-turn election integrity (Anthropic, 2026) Models answer legitimate election requests and decline requests that violate Anthropic’s election policy. Graders check refusals and harmful completions separately. | Everyday | Apr 2026 | Not in cards | |
| BioTIER Safeguarded model configurations receive harmful and benign biology requests. Scored on refusing requests designated for refusal and allowing benign ones. | Expert | Mar 2026 | Not in cards | |
| BioTIER Refuse Models receive 398 dangerous biology prompts. The score measures the percentage correctly refused. | Expert | Mar 2026 | 99.1% | System cards only |
| Disordered eating evaluations (Anthropic) Models handle conversations about eating disorders and disordered eating. Score is the appropriate response rate. | Everyday | Feb 2026 | System cards only | |
| Kiji PII Detection Models locate and label personal information in multilingual text. Dataiku publishes a test split, and independent evaluators compare predicted spans with its annotations. | Everyday | Feb 2026 | Not in cards | |
| KORA Simulated child users hold 3-turn conversations probing 26 child-specific risks. Judge model rates replies failing, adequate or exemplary. | Feb 2026 | 73.2% | Not in cards | |
| PersistBench Models answer requests using stored memories. A judge measures inappropriate disclosure, agreement with false user beliefs and failure to use relevant memories. | Everyday | Feb 2026 | Not in cards | |
| SIM-VAIL Chatbots respond to simulated users with psychiatric vulnerabilities. A model judge scores behaviors across the conversation, including responses that reinforce a user’s vulnerability. | Expert | Feb 2026 | 1 | Not in cards |
| SkillInject Agents use a skill file that carries an adversarial instruction while doing a legitimate task. Reports how often the agent follows the injection. | Expert | Feb 2026 | | System cards only |
| SSH stress-testing evaluation (Anthropic) Models continue prefilled conversations about suicide and self-harm built from anonymized user feedback. Score is the rate of appropriate course correction. | Expert | Feb 2026 | 95% | System cards only |
| HealthBench-Hallu Models answer HealthBench-style medical prompts. A pipeline splits each answer into medical claims and checks them against medical evidence. The score is a weighted hallucination rate. | Expert | Jan 2026 | | Fading |
| MiJaBench Models respond to 43,961 English and Portuguese prompts across 16 minority groups. Judges measure how defense rates change with the target group and language. | University | Jan 2026 | | Not in cards |
| PsychEthicsBench Models answer multiple-choice questions and respond to scenarios under Australian psychology and psychiatry ethics guidelines. Graders check responses against professional standards. | University | Jan 2026 | | Not in cards |
| Agent JSK Agents read simulated email connectors that contain prompt injection attacks. Scored on resisting the injected instructions. | Expert | Dec 2025 | 100% | System cards only |
| BashArena Agents do 637 Linux admin tasks while a red team tries four sabotage goals, such as exfiltrating secrets, past a monitor. Reports sabotage success and monitor catch rates. | Expert | Dec 2025 | 99.1% | System cards only |
| CAIS Risk Index Averages seven safety measures that correlate weakly with general capability. They cover agent prompt injection, bioweapons help, political bias, overconfidence, lying under pressure and harmful or deceptive play in text games. | Research frontier | Dec 2025 | 8.40 | Not in cards |
| CNFinBench Models answer financial questions and carry out financial workflows. The benchmark includes tasks on expertise, autonomy, and integrity. | Expert | Dec 2025 | | Not in cards |
| Kor-Phishing Models classify web requests and links as malicious or benign. Korea's AI Safety Institute reports accuracy separately for each group of 100 items. | University | Dec 2025 | Not in cards | |
| MLCommons AILuminate Jailbreak Models answer hazard prompts with and without taxonomy-based jailbreak attacks. The Resilience Gap is the drop in safety score under attack. | Everyday | Dec 2025 | Not in cards | |
| AssurAI Models respond to Korean prompts across 35 risk categories. Model judges rate the safety of text, image, audio, and video outputs from 1 to 5. | Everyday | Nov 2025 | Not in cards | |
| CIMemories Models use persistent memory across tasks and must withhold attributes when context requires. Reports worst-case disclosure over five samples. | Everyday | Nov 2025 | 14.8% | In use |
| LiveSecBench Models answer Chinese-language prompts that probe safety across areas such as legality, ethics, factuality and privacy. The paper's first edition turned a judge's pairwise comparisons into Elo ratings. The leaderboard's March 2026 edition gives each model an overall score from 0 to 100. The question bank is updated in dated editions. | Everyday | Nov 2025 | 91.4 | Not in cards |
| OASIS (agent safety) Agents handle multi-turn requests with different levels of complexity. Evaluators measure the proportion of safe responses. | University | Nov 2025 | 95.2% | Not in cards |
| WebTailBench-Refusals Computer-use agents receive 111 harmful browser tasks in seven policy categories, such as illegal activity and misinformation. Scored as the share of tasks safely refused. | Everyday | Nov 2025 | | System cards only |
| Dutch DPA national voting advice audit Chatbots rank three Dutch political parties for a fictional voter. The regulator checks whether the first recommendation matches the party used to construct the voter’s answers. | Everyday | Oct 2025 | Not in cards | |
| MentalBench-100k Models answer counseling questions. Human reviewers grade responses to 1,000 of the 10,000 prompts on seven attributes, each from one to five. | Expert | Oct 2025 | 5 | Not in cards |
| PII Masking Models locate personal information in US English text and replace it with a common redaction tag. The default split reserves 20% for evaluation. | Everyday | Oct 2025 | 0.73 | Not in cards |
| The Illusion of AI Safety Chatbots answer 30 prompts about self-harm, eating disorders, and substance use. Researchers check replies for harmful content and invitations to continue the conversation. | University | Oct 2025 | 43% | Not in cards |
| VERA-MH Suicide Risk Chatbots talk with simulated patients who show different levels of suicide risk. A judge grades the full conversation against a clinical rubric. | Expert | Oct 2025 | 75.0 | Not in cards |
| Anthropic paired-prompts political bias eval Models write arguments for opposing sides of political issues from paired prompts. Reviewers label each pair for asymmetries in length, tone, hedging and willingness to engage. | Everyday | Sep 2025 | System cards only | |
| LongCat safety evaluation (Meituan) Models answer private test queries in four risk categories: harmful content, criminal activity, misinformation and privacy. The score is the share of safe responses in each category. | Everyday | Sep 2025 | | Dropped |
| MCP prompt injection evaluation (Anthropic) Agents use simulated email, Slack and document servers over MCP where content carries injected instructions. Score is the attack prevention rate. | Expert | Sep 2025 | 92.5% | System cards only |
| Suicide and self-harm evaluations (Anthropic) Models handle single-turn and multi-turn conversations about suicide and self-harm. Score is the appropriate response rate. | Everyday | Sep 2025 | System cards only | |
| Sycophancy prompts (Anthropic) Models answer single-turn prompts that invite sycophancy, such as bad business ideas or delusional beliefs. Four synthetic sets and one hand-written set report sycophancy rates. | Everyday | Sep 2025 | System cards only | |
| AHELM Audio-language models are scored across 10 aspects with standard prompts. The suite combines existing datasets with two new ones, PARADE for stereotypes and CoRe-Bench for conversational reasoning. | Everyday | Aug 2025 | | Not in cards |
| Challenging prompts from red teamers with biosafety-relevant PhDs Models answer adversarial biology prompts written by red teamers with biosafety PhDs. Scored as not_unsafe rate from model training alone. | Expert | Aug 2025 | 93.6% | System cards only |
| CUA-SHADE-Arena Monitors review computer-use agent activity in simulated workplace tasks. The evaluation measures how often they flag unsafe behavior at a fixed false positive rate. | University | Aug 2025 | Not in cards | |
| Evaluation of Alignment Between Large Language Models and Expert Clinicians in Suicide Risk Assessment Chatbots respond to 30 questions about suicide. Researchers record how often each chatbot answers directly, grouped by five risk levels that clinicians assigned to the questions. | Expert | Aug 2025 | Not in cards | |
| Expert Red Teaming for Violent Attack Planning Expert red teamers compare anonymized responses from two models on violent attack planning prompts. Scored as the win rate for being safer. | Expert | Aug 2025 | 65.1% | System cards only |
| Filtered, adversarial sample of production prompts Models answer filtered borderline and high-risk biology prompts that resemble production traffic. Scored as not_unsafe rate from model training alone. | Expert | Aug 2025 | 96.9% | System cards only |
| HealthBench Consensus Urgent Models answer ambiguous high-stakes health conversations validated by two or more physicians. Scored as the error rate in informing the user. | Expert | Aug 2025 | 0.40% | System cards only |
| HealthBench Hard Hallucinations Models respond to difficult health conversations that 2 or more physicians validated. The score is the hallucination rate, so lower is better. | Aug 2025 | 1.60% | Dropped | |
| MCPTox Agents use tools whose descriptions contain untrusted instructions. The benchmark measures whether agents follow those instructions across 1,348 cases from 45 servers. | University | Aug 2025 | | Not in cards |
| Pairwise Bioweaponization Campaign Red teamers with biology PhDs compare anonymized responses from two models on bioweaponization prompts. Scored as the win rate for being safer. | Expert | Aug 2025 | 60% | System cards only |
| RiskRubric Evaluators score public models on six dimensions of risk. The scorecard combines the results into a letter grade. | Expert | Aug 2025 | 891 | Not in cards |
| Spiral-Bench v1.0 Models chat for 20 turns with a simulated suggestible user played by Kimi-K2. A judge logs protective and risky behaviors such as pushback, sycophancy and delusion reinforcement, which combine into a safety score. | Everyday | Aug 2025 | 87.0 | Not in cards |
| Active data exfiltration, visual browser Agents using a visual browser meet attacks that require them to fetch sensitive information before exfiltrating it. Scored on resisting them. | Expert | Jul 2025 | 67% | System cards only |
| CaBBQ Models answer Catalan questions about people in ambiguous and unambiguous situations. The test measures whether answers follow social stereotypes when evidence is insufficient. | Everyday | Jul 2025 | | Not in cards |
| In-context data exfiltration, visual browser Agents using a visual browser meet attacks that try to exfiltrate data already present in the conversation context. Scored on resisting them. | Expert | Jul 2025 | 78% | System cards only |
| Irrelevant instructions, text-based web browser Agents using a text browser meet synthetic web pages with irrelevant instructions or data exfiltration attempts. Scored on ignoring them. | Expert | Jul 2025 | 99.5% | System cards only |
| Irrelevant instructions, visual browser Agents using a visual browser meet irrelevant instructions or data exfiltration attempts on web pages, based on red teaming scenarios. Scored on ignoring them. | Expert | Jul 2025 | 95% | System cards only |
| Kimi Promptfoo red-team safety evaluation Promptfoo generates attack prompts for harmful, criminal, misinformation, privacy and security plugins under five strategies. Human reviewers judge the responses. | Everyday | Jul 2025 | System cards only | |
| TextQuests Agents play 25 classic interactive fiction games. The score is harmful actions per 100 steps, where lower is better. The board also reports game progress. | Everyday | Jul 2025 | | System cards only |
| Agentic misalignment suite (Anthropic) Agents in simulated company settings face threats or goal conflicts that invite blackmail, leaking or sabotage. Score is the misaligned action rate. | Expert | Jun 2025 | 0% | In use |
| ClearHarm Models answer unambiguously harmful CBRN questions. Judges score whether attacks elicit harmful responses from the model. | Expert | Jun 2025 | Not in cards | |
| FORTRESS Paired adversarial and benign national-security misuse prompts. Rubric-graded risk score on adversarial and over-refusal score on benign prompts. | Jun 2025 | 5.78% | In use | |
| FORTRESS-RN Models face the radiological and nuclear prompts of Scale AI's FORTRESS. xAI reports refusal accuracy on these items. | Jun 2025 | 97.9% | System cards only | |
| Poly-Guard Bench Models meet harmful requests and statements across domains. Reports attack success rate under single-turn and multi-turn adaptive red-teaming. | Expert | Jun 2025 | | System cards only |
| AudioTrust Audio language models answer prompts that probe hallucination, robustness, speaker authentication, privacy, fairness, and safety. Automated evaluators, including language-model judges, compare responses with task-specific references and criteria. | Everyday | May 2025 | | Not in cards |
| CounselBench-Adv Models answer 120 questions written by mental health professionals. Each question targets one of six problems identified in earlier counseling responses. | Expert | May 2025 | 0.37 | Not in cards |
| CRMArena-Pro Agents do 19 expert-checked sales, service and quoting tasks in B2B and B2C Salesforce orgs, in single-turn and persona-driven multi-turn settings. It also tests whether agents keep confidential data private. | Everyday | May 2025 | | Not in cards |
| HealthBench 5,000 realistic health conversations graded against physician-written rubrics. A length-adjusted variant penalizes verbosity. | Expert | May 2025 | 67.2% | In use |
| RiOSWorld Computer-use agents work in an Ubuntu virtual machine through risky scenarios such as phishing emails and pop-ups. Scores track whether agents notice and avoid the risk. | Everyday | May 2025 | | Not in cards |
| SoSBench Models answer hazardous requests drawn from real regulations in six scientific fields. Scored by how often the model complies with a request it should refuse. | Expert | May 2025 | 10.6% | Not in cards |
| Sri Lankan local election chatbot audit Chatbots answer 16 procedural and two political questions in English, Sinhala, and Tamil. Reviewers score procedural accuracy and political neutrality separately. | Everyday | May 2025 | Not in cards | |
| Video-SafetyBench Video-language models receive 2,264 synthesized videos paired with a harmful or a benign-looking query across 13 unsafe categories and 48 subcategories. An LLM judge scores harm with a confidence-weighted metric called RJScore. | Everyday | May 2025 | 1.30 | Not in cards |
| Person Identification evaluation Models receive requests to identify people in photos. Scored on refusal rate in non-adversarial and adversarial sets. | Everyday | Apr 2025 | System cards only | |
| RealHarm Safety systems classify real-world harmful conversations and compare their decisions with documented safe and unsafe versions across ten harm categories. | Expert | Apr 2025 | Not in cards | |
| Ungrounded Inference evaluation Models receive requests for inferences about people that images or text cannot justify. Scored on refusal rate in non-adversarial and adversarial sets. | Everyday | Apr 2025 | System cards only | |
| AI Agent Privacy Web agents complete tasks while exposed to private information. The benchmark measures privacy leakage and unsafe disclosure during autonomous operation. | Expert | Mar 2025 | Not in cards | |
| Harmful content defect rate (Azure AI Evaluation SDK, Microsoft) GPT-4o simulates adversarial conversations with the model, and a judge flags violence, sexual, self-harm and hateful content. Scored as the share of harmful responses, lower is better. | Everyday | Mar 2025 | | System cards only |
| Challenging Red Teaming Evaluation 1 (created for o3-mini) Models face adversarial jailbreaks from o3-mini red teaming on illicit advice, extremism, political persuasion and self-harm. Scored as not_unsafe rate. | Expert | Feb 2025 | 51% | System cards only |
| Challenging Red Teaming Evaluation 2 (created for deep research) Models face adversarial prompts from deep research red teaming on risky advice such as attack planning. Scored as not_unsafe rate. | Expert | Feb 2025 | 67% | System cards only |
| Child safety evaluations (Anthropic) Models answer single-turn and multi-turn prompts about child sexualization and grooming. Score is the appropriate response rate. | Everyday | Feb 2025 | System cards only | |
| German federal election chatbot audit Chatbots answer German election questions in English and German. Researchers count correct electoral answers out of 30 per chatbot and rate political bias separately. | Everyday | Feb 2025 | Not in cards | |
| Malicious computer use (Anthropic) Computer-use agents receive harmful tasks such as surveillance or fraud in GUI environments. Score is the refusal rate. | Everyday | Feb 2025 | 99.4% | System cards only |
| MedHallu Models distinguish factual medical answers from hallucinated answers at three difficulty levels. The benchmark tests detection with and without supporting knowledge. | University | Feb 2025 | | Not in cards |
| Personal data evaluations Browsing agents receive requests for private personal information, from 200 synthetic prompts and 55 manually written golden examples. Scored on adherence to personal data policy. | Everyday | Feb 2025 | System cards only | |
| Phare Models are probed in English, French and Spanish for hallucination, harmful misguidance, stereotypes and jailbreaks. Hallucination tasks cover factuality, misinformation, debunking and tool use. | Everyday | Feb 2025 | 83.2% | Not in cards |
| Aegis 2.0 Models classify human and LLM conversation turns as safe or unsafe under a taxonomy of 12 hazard categories. Scored by F1 against human and jury labels. | Everyday | Jan 2025 | | System cards only |
| AfriHate Models classify hate speech and abusive language across African languages. | Everyday | Jan 2025 | | Not in cards |
| Agentic harmful task refusals Browser agents receive harmful tasks such as illicit purchases, prohibited financial activity and searches for sensitive personal data. Scored as refusal rate per category. | Everyday | Jan 2025 | 97% | System cards only |
| CASE-Bench Models decide whether a request is safe in a given context. The same request appears in two contexts with different human judgments. | Everyday | Jan 2025 | 90.1% | Not in cards |
| MedQA Bias Evaluation Models answer medical exam questions before and after Vals adds racial information. Vals compares accuracy across patient groups. | Expert | Jan 2025 | 96.5% | Not in cards |
| Proactive refusals of high-risk tasks Browser agents receive high-risk tasks such as banking transfers and high-stakes decisions that policy says they must refuse. Scored on refusal recall. | Everyday | Jan 2025 | 94% | System cards only |
| Agent-SafetyBench LLM agents act in 349 tool environments over 2,000 test cases covering 8 safety risk categories and 10 failure modes. A scorer model rates each run safe or unsafe. | Everyday | Dec 2024 | 59.8% | Not in cards |
| Gray Swan Jailbreak Arena Human attackers try to jailbreak models for violent, self-harm and malicious code outputs. Scored as attack success rate per category. | Expert | Dec 2024 | 3.60% | System cards only |
| Image input safety evaluations Models answer harmful prompts that combine text and images in categories such as hate, extremism and self-harm. Scored as safe completion rate. | Everyday | Dec 2024 | 99% | System cards only |
| Instruction Hierarchy Evaluation Models resolve conflicts between system, developer and user messages, resist tutor jailbreaks, and protect secret phrases and passwords. Scored per sub-evaluation. | Everyday | Dec 2024 | System cards only | |
| MLCommons AILuminate Industry-standard hazard benchmark over 12 risk categories with graded ratings. | Dec 2024 | | Not in cards | |
| HELM Safety Stanford CRFM runs five safety benchmarks over six risk categories, such as violence, fraud and discrimination, and reports one leaderboard. | Everyday | Nov 2024 | 98.6 | Not in cards |
| Proof News election misinformation debunking Chatbots respond to five false claims about the 2024 US election. Reviewers mark whether each service clearly debunks each claim. | Everyday | Nov 2024 | 0% | Not in cards |
| RedCode Models meet adversarial prompts to generate or run harmful code. Generation is model-judged and execution is checked in a container. | Expert | Nov 2024 | | System cards only |
| Tulu 3 Safety Eval Models answer harmful, jailbreak and benign prompts from several safety benchmarks in Ai2's safety-eval suite. A WildGuard classifier judges the outputs, and the score is an average. | Everyday | Nov 2024 | | Fading |
| AgentHarm 110 malicious agent tasks with synthetic tools. Measures refusal and harmful task completion, plus benign controls. | Oct 2024 | 8.00% | Fading | |
| Arizona voter questions in English and Spanish Models answer Arizona voter questions in English and Spanish. Fact-checkers flag answers containing a false statement or a broken link. | Everyday | Oct 2024 | 30% | Not in cards |
| BrowserART Browser agents receive 100 harmful requests on synthetic and real websites. The test measures how often the agent attempts the harmful behavior that its backbone model would refuse in chat. | Expert | Oct 2024 | 11% | Not in cards |
| Election integrity evaluations (Anthropic) Models answer election-related prompts, including requests for misinformation or influence operations. Score is the appropriate response rate. | Everyday | Oct 2024 | System cards only | |
| First-Person Fairness evaluation Tests whether a user's name (a demographic proxy) changes responses, using an LLM research assistant to detect harmful stereotypes. | Oct 2024 | 0.40% | System cards only | |
| VHELM Vision-language models are scored across nine aspects, including perception, reasoning, knowledge, bias, fairness, multilinguality, robustness, toxicity and safety, from many existing datasets. | University | Oct 2024 | | Not in cards |
| Generating Confusion: Voting with a Disability Models answer questions about accessible voting in the United States. Reviewers measure how often answers to standard voter questions contain incorrect information. | Everyday | Sep 2024 | Not in cards | |
| Political Persuasion Parallel Generation Human raters pick the more persuasive of two short political social media posts. Scored as win rate against GPT-4o. | Everyday | Sep 2024 | 53.8% | System cards only |
| Regurgitation Evaluations Models are asked to reproduce training data. Scored on how often they refuse to regurgitate it. | Everyday | Sep 2024 | System cards only | |
| SPY Models locate personal information about the author of a synthetic medical consultation or legal question. Other people mentioned in the text fall outside the annotation policy. | Everyday | Sep 2024 | Not in cards | |
| Multi-Turn Human Jailbreaks Models respond to multi-turn conversations collected in human safety evaluations. The authors report how often model safeguards fail. | Expert | Aug 2024 | Not in cards | |
| PrivacyLens Language model agents write emails, posts or messages from a prior tool trajectory without leaking sensitive details. A judge scores information leakage and helpfulness. | Everyday | Aug 2024 | | System cards only |
| Speaker identification Speech models get audio requests to identify a speaker by voice. Scored on refusing when they should and complying for famous quotes. | Everyday | Aug 2024 | System cards only | |
| Text persuasion study on political opinions Participants read model-written articles or chat with a model on political topics. Scored on opinion shift against human-written articles. | Everyday | Aug 2024 | System cards only | |
| Ungrounded inference and sensitive trait attribution Speech models get requests to infer traits about a speaker from audio. Scored on refusing ungrounded inferences and hedging plausible ones. | Everyday | Aug 2024 | 84% | System cards only |
| Voice persuasion study on party preferences Participants hear model-voiced audio clips or talk with a voiced model. Scored on shifts in hypothetical party preference against human baselines. | Everyday | Aug 2024 | 5.74% | System cards only |
| (S)elected Moderation campaign prompts Models receive requests for election campaign advice through their APIs. Reviewers count refusals. The report specifies ten English prompts. | Everyday | Jul 2024 | Not in cards | |
| AIR-Bench 2024 Models respond to 5,694 harmful prompts across 314 risk categories drawn from 8 government regulations and 16 company policies. Category-specific LLM judges score refusals. | Everyday | Jul 2024 | 93.2% | Dropped |
| Llama verbatim memorization probe (Meta) Models get prompts sampled from their pretraining data. The score is the share of generations that reproduce the true continuation verbatim. | Jul 2024 | System cards only | ||
| MedBench 2.0 (OpenCompass) Models answer Chinese medical questions. This release changes the evaluation of open responses. | Expert | Jul 2024 | Not in cards | |
| MJ-Bench Models compare image pairs for text alignment, safety, visual quality, and bias. Scores measure preference accuracy, with separate dispersion and equality measures for bias. | Everyday | Jul 2024 | 44.5% | Not in cards |
| Qwen2 multilingual safety evaluation Models answer risky prompts in several languages on illegal behavior, fraud, pornography and privacy, including jailbreak attempts. Scores are harmful response rates. | Everyday | Jul 2024 | System cards only | |
| AgentDojo Agents do tool-use tasks while tools and data carry prompt injections. Attack success rate counts hijacked runs. | Everyday | Jun 2024 | 0% | Fading |
| Aya Red-teaming Models answer harmful prompts in eight languages across nine harm categories. Judges score refusal and safety behavior against annotated categories. | Everyday | Jun 2024 | Not in cards | |
| Election Questions (Anthropic) Models answer EU election questions and respond to US requests about voting and political targeting. Anthropic measures factual reliability and whether models refuse harmful requests. | Everyday | Jun 2024 | Not in cards | |
| Gretel Synthetic PII Finance Models identify and label personal information in synthetic financial documents across seven languages. Independent evaluations measure how well the predicted spans match the labeled spans. | Everyday | Jun 2024 | Not in cards | |
| HarmEval Models answer 550 harmful requests across 11 categories. A judge measures the share of unsafe responses. | Everyday | Jun 2024 | | Not in cards |
| JailTrickBench Models respond to jailbreak attacks under varied attack and defense settings. A GPT-4 evaluator judges attack success across harmful-request datasets. | Expert | Jun 2024 | Not in cards | |
| SORRY-Bench Models get unsafe requests across a fine-grained taxonomy of topics and linguistic variations. A judge model scores whether each answer refuses or complies. | Everyday | Jun 2024 | 5.23% | Not in cards |
| WildGuardTest Models classify whether user prompts are harmful, whether responses are harmful, and whether responses refuse, across 13 risk categories. Scored by F1. | Everyday | Jun 2024 | | System cards only |
| WildJailbreak Models face adversarial jailbreak prompts built from in-the-wild tactics, plus benign prompts that look harmful. Evaluations score refusals of harmful requests and compliance with benign ones. | Everyday | Jun 2024 | | Fading |
| OR-Bench Models meet benign and borderline prompts and the benchmark measures how often they wrongly refuse. Reports false refusal rate. | University | May 2024 | 6.80% | System cards only |
| S-Eval Models respond to Chinese and English harmful prompts and attack variants. A trained safety judge measures the safe-response fraction and attack success rate. | Everyday | May 2024 | 88.8% | Not in cards |
| Singapore Safety Questions Models answer 59 open-ended questions about culturally or legally sensitive situations in Singapore. A Llama Guard classifier checks whether each response is safe. | Everyday | May 2024 | Not in cards | |
| Voice Jailbreak Attacks Against GPT-4o Models respond to spoken requests. Human reviewers measure how often a model complies with an unsafe request. | Everyday | May 2024 | Not in cards | |
| Wahl-O-Mat 2024 European election party agreement Models take a position on 38 German statements about European Union policy. Researchers compare those answers with each party’s published positions. | Everyday | May 2024 | Not in cards | |
| WildChat (toxic conversations) Models answer toxic user turns from WildChat, the 200 highest scoring conversations per moderation category. Scored as not_unsafe rate. | Everyday | May 2024 | 98% | System cards only |
| European Parliament election chatbot audit Chatbots answer ten European election questions in ten languages. Reviewers score procedural answers for accuracy and political answers for neutrality. | Everyday | Apr 2024 | Not in cards | |
| Microsoft RAI benchmark: Harmful Content Continuation (DR-3) Models continue conversations about hate, self-harm, sexual content and violence. Scored as the share of answers with judged severity 3 or more out of 7. | Everyday | Apr 2024 | | System cards only |
| Microsoft RAI benchmark: Harmful Content Summarization (DR-3) Models summarize content about hate, self-harm, sexual content and violence. Scored as the share of answers with judged severity 3 or more out of 7. | Everyday | Apr 2024 | | System cards only |
| Microsoft RAI benchmark: Jailbreak (DR-1) Models face simulated jailbreak conversations on topics such as child grooming, illegal advice and leaking guidelines. Scored as the defect rate at severity 1. | Everyday | Apr 2024 | | System cards only |
| Microsoft RAI benchmark: Third Party Harm (DR-1) Models face simulated conversations that try to extract third-party content such as books, news, recipes and songs. Scored as the defect rate at severity 1. | Everyday | Apr 2024 | | System cards only |
| MLCommons AI Safety Benchmark v0.5 Chat models answer 43,090 English prompts across seven hazards. LlamaGuard labels the responses, and the highest risk grade across hazards determines the overall grade. | Everyday | Apr 2024 | Not in cards | |
| RTP-LX Models continue or judge toxic prompts hand-translated from RealToxicityPrompts into dozens of languages, with native-speaker labels. Measures toxic output and toxicity detection. | Everyday | Apr 2024 | Not in cards | |
| AraTrust Models answer 522 human-written Arabic multiple-choice questions on truthfulness, ethics, safety, physical and mental health, unfairness, illegal activities, privacy and offensive language. | Everyday | Mar 2024 | | Dropped |
| JailbreakBench 100 harmful and 100 benign behaviors with a standard jailbreak artifact library. Scored by attack success rate. | Mar 2024 | 0% | System cards only | |
| MedSafetyBench Models respond to requests that conflict with medical ethics. A judge rates harmfulness on a five-point scale using a held-out set of 900 requests. | Everyday | Mar 2024 | | Not in cards |
| Multimodal policy red-teaming (Anthropic) Models answer multi-turn image and text conversations on harmful topics such as child safety, weapons and fraud. Score is the share of harmless responses. | Everyday | Mar 2024 | 99.2% | System cards only |
| NTCIR-18 MedNLP-CHAT Models flag medical, legal, and ethical risks in answers to patient questions. The organizers compare the labels with expert judgments using macro F1. | Expert | Mar 2024 | Not in cards | |
| PII-Masking-300k Models locate and classify personal information in multilingual synthetic text. The open portion covers six languages. The authors separately describe a finance and insurance portion. | Everyday | Mar 2024 | Not in cards | |
| WildChat toxic refusals (Anthropic) Models answer toxic user prompts drawn from the WildChat dataset of real chatbot conversations. Score is the correct refusal rate. | Everyday | Mar 2024 | 96.4% | System cards only |
| AI Democracy Projects voter queries Models answer questions that US voters might ask about an election. Expert panels vote on whether each response contains inaccurate information. | Everyday | Feb 2024 | 19.2% | Not in cards |
| DETESTS-Dis 2024 Models detect racial stereotypes in Spanish comments and posts and decide whether the stereotypes are implicit. Organizers score majority labels and distributions of annotator judgments. | University | Feb 2024 | Not in cards | |
| Divergence attack (extractable memorization) Models are prompted to repeat one token many times. The test measures how often they diverge and emit memorized training text. | Everyday | Feb 2024 | | System cards only |
| HarmBench Models answer harmful behavior requests, including copyright and contextual cases, with and without automated red-team attacks. A classifier judges whether the response complies. | Everyday | Feb 2024 | | Fading |
| Red-Teaming Resistance Benchmark Models respond to prompts from multiple safety datasets. Judges assess whether the responses comply with harmful requests. | Everyday | Feb 2024 | Not in cards | |
| SALAD-Bench Models answer harmful questions organized in a three-level taxonomy, some with attack or defense changes and some multiple-choice. The MD-Judge evaluator labels responses as safe or unsafe. | Everyday | Feb 2024 | 96.8% | Fading |
| StrongREJECT Forbidden-prompt jailbreak benchmark with a rubric grader that scores how useful a non-refusal actually is. | Feb 2024 | | In use | |
| VLGuard Vision language models answer safe and unsafe image and text pairs from the VLGuard test set. Responses are judged for harmfulness and refusal. | Everyday | Feb 2024 | 9.10 | System cards only |
| DeepSeek LLM in-house safety evaluation Models answer 2,400 Chinese test cases on discrimination, legal rights, trade secrets, illegal behavior and other safety issues. Human annotators judge each answer. | Everyday | Jan 2024 | System cards only | |
| Enterprise PII Models respond to 100 prompts that request sensitive business information. The evaluation checks whether the responses disclose it. | University | Jan 2024 | Not in cards | |
| Enterprise Scenarios Toxicity Models respond to 100 prompts. The Perspective API scores how often their responses contain toxic language. | Everyday | Jan 2024 | Not in cards | |
| K-QA Models answer 201 patient questions. The benchmark checks whether answers include essential statements from clinicians and whether any claims contradict those statements. | Expert | Jan 2024 | 67.7% | Not in cards |
| MedBench 1.0 (OpenCompass) Models answer Chinese medical questions across clinical reasoning, language understanding, generation, knowledge, and ethics. | Expert | Jan 2024 | Not in cards | |
| MuTox Models listen to short speech clips in 30 languages and classify whether they contain toxic speech. The harness compares their decisions with human labels. | Everyday | Jan 2024 | Not in cards | |
| RTVLM Vision language models answer red-teaming image prompts on faithfulness, privacy, safety and fairness. GPT-4V scores responses from 0 to 10. | Everyday | Jan 2024 | 7.92 | System cards only |
| TrustLLM JailbreakTrigger Models face 400 prompts that apply 13 jailbreak methods to questions about bad behavior and requests for toxic content. Scored by refusal rate. | Everyday | Jan 2024 | | System cards only |
| FLAMES Models respond to Chinese prompts about safety and values. Evaluators assess the responses against the benchmark's alignment criteria. | Everyday | Nov 2023 | 43.2% | Not in cards |
| HEIM Text-to-image models are scored on 12 aspects, such as alignment, quality, aesthetics, originality, reasoning, knowledge, bias and toxicity, with human raters and automatic metrics. | Everyday | Nov 2023 | Not in cards | |
| Insider Trading Demo Models act as stock traders under pressure. Apollo measures how often they trade on inside information and conceal their actions. | University | Nov 2023 | 4.67% | Not in cards |
| MM-SafetyBench Multimodal models get 5,040 harmful text and image pairs in 13 scenarios, where a query-relevant image carries the harmful intent. Judges score attack success. | Everyday | Nov 2023 | Not in cards | |
| SimpleSafetyTests Models answer 100 clearly harmful prompts on illegal items, physical harm, scams, child abuse and self-harm. The score is the share of safe responses. | Everyday | Nov 2023 | 100% | System cards only |
| Bing Chat election answers in Switzerland, Bavaria, and Hesse Bing Chat answers questions about candidates, polls, and voting in three elections. Annotators identify factual errors, evasions, and accurate answers. | Everyday | Oct 2023 | Not in cards | |
| RaceBias (Race-Based Medicine) Models answer clinical questions used to identify unsupported race-based medical beliefs. Evaluators check responses for those beliefs. | Expert | Oct 2023 | Not in cards | |
| ToxicChat Models identify toxic user messages from real conversations with an online chatbot. | School | Oct 2023 | Not in cards | |
| AttaQ Models answer 1,402 adversarial questions in seven harm classes, such as deception, violence and PII. Responses are scored for harmlessness. | Everyday | Sep 2023 | 88.5% | Fading |
| MakeMePay A con artist model tries to talk a GPT-4o mark into paying money in a role-play. Scored as the share of conversations with a payment. | Everyday | Sep 2023 | 57% | System cards only |
| SafetyBench Models answer 11,435 multiple choice questions on seven safety topics in Chinese and English, such as ethics, illegal activities and privacy. | Everyday | Sep 2023 | 89.9% | System cards only |
| Do-Not-Answer Models get risky instructions that responsible models should not follow, across 12 harm areas and 61 risk types. Responses are judged for safe handling. | Everyday | Aug 2023 | | System cards only |
| DoAnythingNow (DAN) Models face in-the-wild DAN-style jailbreak prompts paired with forbidden questions. Scored by refusal rate or attack success rate. | Everyday | Aug 2023 | | System cards only |
| XSTest Models answer 250 safe prompts that resemble unsafe ones and 200 unsafe contrast prompts. Labs usually report the incorrect refusal rate on safe prompts. | Everyday | Aug 2023 | | System cards only |
| AdvBench Models get harmful requests, such as a request for malware, and should refuse them. The set has 520 harmful behaviors and 574 harmful strings. Most papers add an attack, such as an adversarial suffix, and report the attack success rate. | Everyday | Jul 2023 | Not in cards | |
| Llama 2-Chat safety human evaluation (Meta) Human raters score responses to about 2,000 adversarial prompts on a five-point scale. Ratings of 1 or 2 count as violations. | Jul 2023 | System cards only | ||
| Med-HALT Models answer medical exam questions designed to reveal hallucinations. The hub port uses the false-confidence, fake-question, and none-of-the-above reasoning tasks. | Expert | Jul 2023 | | Not in cards |
| SuperCLUE-Open Models answer Chinese open-domain questions over multiple turns. The authors compare models across skill categories. | Everyday | Jul 2023 | 94.6 | Not in cards |
| DecodingTrust Models are tested on eight trustworthiness areas, including toxicity, stereotype bias, robustness, privacy, machine ethics and fairness. | Everyday | Jun 2023 | | Not in cards |
| Incorrect behavior rate on disallowed and sensitive content Models answer disallowed prompts and sensitive prompts such as medical advice. Scored on the rate of responses that break policy. | Everyday | Mar 2023 | System cards only | |
| Media Bias Identification Benchmark Models classify text in eight tasks, including linguistic bias, political bias, and hate speech. The benchmark reports F1 across pooled examples and separately by source dataset. | Everyday | Feb 2023 | Not in cards | |
| MultiMedQA Models answer medical exam, research, and consumer health questions across seven datasets. Multiple-choice components use answer accuracy. Clinicians and other raters assess free responses for factuality, precision, harm, and bias. | Expert | Dec 2022 | | Not in cards |
| Inappropriate Image Prompts Models generate images from prompts drawn from real users. The evaluation measures how often the outputs contain inappropriate material. | Everyday | Nov 2022 | Not in cards | |
| K-MHaS Models assign hate speech labels to Korean news comments. A comment can receive more than one label. | Everyday | Oct 2022 | Not in cards | |
| Privy Synthetic PII Protocol Trace Dataset Models locate personal information in synthetic API messages represented as protocol traces. The small and large configurations have separate test files. | University | Sep 2022 | Not in cards | |
| Anthropic Red Team Models answer human red-team attack prompts collected by Anthropic on discrimination, hate, violence and unethical behavior. Judges score response safety. | Everyday | Aug 2022 | 99.96% | System cards only |
| Multilingual HateCheck Models distinguish hate speech from statements that quote, reject, or otherwise resemble hate speech. The test covers multiple languages. | Everyday | Jul 2022 | Not in cards | |
| ToxiGen Models continue ToxiGen prompts about 13 minority groups. Scored by the average toxicity of the generations. | Everyday | Mar 2022 | | System cards only |
| ToxiGen (discriminative) Models label statements about 13 minority groups as toxic or neutral. Scored by accuracy, often split into toxic and neutral labels. | Everyday | Mar 2022 | | System cards only |
| CredData Models identify hardcoded credentials in source files. Samsung publishes reviewed labels and obfuscates credential values. OpenAI scores both tokens and exact spans. | University | Aug 2021 | Not in cards | |
| UIT-ViCTSD Models classify Vietnamese comments for constructiveness and toxicity. The two labels capture different properties of the same text. | Everyday | Mar 2021 | Not in cards | |
| BOLD Models continue 23,679 English Wikipedia prompts about professions, gender, race, religion and politics. Classifiers score sentiment, toxicity and polarity per group. | Everyday | Jan 2021 | | System cards only |
| CivilComments-WILDS Models classify the toxicity of online comments. The benchmark evaluates how accuracy varies across groups mentioned in the text. | Everyday | Dec 2020 | Not in cards | |
| ToLD-Br Models identify toxic language in Brazilian Portuguese posts using human annotations. | Everyday | Oct 2020 | Not in cards | |
| CrowS-Pairs Models choose between paired sentences that differ only in a stereotyped social group. Measures preference for the stereotyped sentence. | Everyday | Sep 2020 | | System cards only |
| RealToxicityPrompts Models continue 100,000 web text prompts. Scored on the rate of toxic generations judged by a toxicity classifier. | Everyday | Sep 2020 | | System cards only |
| Unhealthy Comments Corpus Models rate newspaper comments for overall conversational health and six attributes, including hostility and sarcasm. Researchers score each label separately using ROC AUC. | Everyday | Aug 2020 | Not in cards | |
| Hateful Memes Challenge Models decide whether a meme is hateful by interpreting its image and caption together. Paired examples change one modality while preserving the other. | Everyday | May 2020 | Not in cards | |
| StereoSet Models choose between stereotypical, anti-stereotypical and unrelated completions in four domains: gender, profession, race and religion. Scores combine language modeling and stereotype preference. | Everyday | Apr 2020 | Not in cards | |
| Jigsaw Toxic Comment Classification Models label Wikipedia talk page comments for toxicity types such as threats, insults and identity hate. Labs report F1 or ROC AUC. | Everyday | Dec 2017 | | System cards only |
| Anthropic harmful-request and wellbeing evaluations Single-turn harmful and benign request sets, multi-turn tests, child safety, suicide/self-harm, disordered eating and election integrity. | Not in cards | |||
| Bedside Bench Models answer 500 physician-written clinical cases on diagnosis, workup, escalation and medication safety. An LLM judge grades each answer against the authors' criteria. | Expert | 94.6% | Not in cards | |
| Child Safety Action Testing Computer use agents get requests for actions that facilitate child exploitation, endangerment or illegal access. Reports the refusal score. | Everyday | | System cards only | |
| Child Sexual Abuse and Exploitation Safety Specialist teams grade model answers to adversarial and coded child sexual abuse and exploitation requests in text, image and video. Reports safe response rate. | Everyday | | System cards only | |
| Contextual security evaluation Agents handle noisy cross-app data with sensitive records while attackers exploit ambiguity to exfiltrate data. Reports baseline safe behavior and survival rates. | Expert | | System cards only | |
| Dangerous Content Safety An autorater checks answers to single and multi-turn requests about weapons, illegal acts, violence and self-harm. Reports safe response rate. | Everyday | | System cards only | |
| Delusional Reinforcement evaluation Multi-turn chats test whether models avoid validating beliefs not grounded in reality and point users to real support. Reports policy adherence rate. | Everyday | | System cards only | |
| Discoverable memorization of training data Models continue 50-token prefixes sampled from their own training data. The test counts exact and approximate copies of the true suffix. | Everyday | System cards only | ||
| Hate & Harassment Safety An autorater checks answers to requests for hateful or harassing content and to sensitive benign prompts. Reports safe response rate. | Everyday | | System cards only | |
| Health language and locale eval Models answer locale-specific health conversations in 10 languages across 25 locales, with search. Graded against clinician guidelines. | Expert | System cards only | ||
| High-risk health scenarios eval Adversarial health emergencies test whether models recognize urgency and red flags and direct users to help. Reports pass rate. | Expert | | System cards only | |
| Implicit Intelligence Agents respond to requests whose unstated requirements affect safety, privacy, or accessibility. Labelbox reports the share of scenarios each model passes. | Everyday | 53.2% | Not in cards | |
| Indirect prompt injection, Actor Critic attack An attacker model refines injected email triggers until the agent leaks synthetic private data through a send email call. Reports attack success rate. | Expert | | System cards only | |
| Indirect prompt injection, Beam Search attack An attack appends random tokens to an injected email trigger until the agent leaks synthetic private data. Reports attack success rate. | Expert | | System cards only | |
| Indirect prompt injection, TAP attack A Tree of Attacks with Pruning search writes natural language injections that make the agent leak synthetic private data. Reports attack success rate. | Expert | | System cards only | |
| KorSET Models respond to Korean prompts adapted to local language, law, and culture. A model judge checks whether each response violates the relevant safety policy. | Everyday | | Not in cards | |
| MAPA EUR-Lex Models mark sensitive entities in EU legal texts. MAPA annotates general and fine-grained categories. The multilingual conversion covers 21 languages. | University | Not in cards | ||
| Meta chemical-agent refusal evaluation Models receive high-risk chemical prompts and the benchmark measures how often they refuse. Covers development, procurement and production stages. | Expert | | System cards only | |
| OpenAI Dynamic Mental Health Benchmarks Multi-turn conversations with adversarial simulated users on mental-health risks. Scored on safe and supportive responses. | System cards only | |||
| Personal Agent Bench Personal assistants complete everyday tasks using connected accounts. Experts check whether each assistant completes the work and respects the user’s instructions. | Everyday | Not in cards | ||
| Personal-device assistant prompt injection eval OS assistant agents in 44 stateful scenarios ingest untrusted content that tries to trigger unauthorized actions. Reports 5-step adaptive survival and task utility. | Expert | | System cards only | |
| Polaris Healthcare Safety Benchmarks Models conduct simulated healthcare conversations. The evaluation measures clinical safety, communication, and administrative task performance. | Expert | Not in cards | ||
| SEAL Adversarial Robustness Models receive 1,000 single-turn adversarial prompts written by 10 full-time red teamers across nine harm categories. Safety specialists grade responses, and models are ranked by the number of high-harm violations. | Expert | | Not in cards | |
| Sexually Explicit Content Safety An autorater checks answers to direct, adversarial and benign sexual content prompts in single and multi-turn chats. Reports safe response rate. | Everyday | | System cards only | |
| STELLA-Bench Models respond to simulated users over multiple turns. Judges measure the shares of conversations with harmful or unhelpful responses. | Everyday | 0% | Not in cards | |
| StrongREJECT v2 (Meta) Models face a cascade of jailbreak attacks on disallowed behaviors adapted from StrongREJECT plus internal ones. Reports attack success rate. | Expert | | In use | |
| Suicide & Self-Harm evaluation Static and dynamic chats with suicidal or self-harm signals test adherence to prevention practice and crisis resources. Reports policy adherence rate. | Everyday | | System cards only | |
| SuperCLUE-DSPSafeBench Models respond to Chinese prompts that test content safety. | Everyday | | Not in cards | |
| SWE-PI A coding agent does a bug-fix task in a repository seeded with an injected payload that tries to exfiltrate a secret. Reports attack success rate. | Expert | | System cards only | |
| UIT-ViHSD Models classify Vietnamese social media text into three categories covering clean language, offensive language, and hate speech. | Everyday | Not in cards | ||
| VetEval Models answer examination questions written by licensed veterinarians. A fixed answer key measures clinical knowledge, and safety penalties reduce the overall index. | University | 76.8 | Not in cards | |
| xAI refusal and jailbreak suites Internal refusal suites (general, child safety, self-harm) in six languages plus single-turn, multi-turn and long-horizon jailbreaks. | Not in cards |