Gram (Gauging Realistic Agentic Misbehavior)
Automated auditing over 17 simulated deployment scenarios that incentivize sabotage by coding and research agents. LLM judges plus human review.
Builds on Petri. No public repo found.
- Made by
- Google DeepMind
- Released
- May 2026
- Kind
- Deception and alignment
- Scored
- Graded by a model
- Answers
- Agent task
- Strongest model tested
-
Gemini 4 Argon #9 on the capability index - Last in a system card
- October 2026
- Last model added
- October 2026
In the authors’ words
We introduce Gram, an automated alignment auditing framework to assess the propensity of AI agents to engage in sabotage. We evaluate Gemini models across 17 simulated agentic deployment scenarios that incentivize sabotage. We find Gemini models misbehave in about 2-3% of our simulated trajectories. Many of these cases are explained by "overeagerness" in Gemini models resulting in both excessive role-playing and goal-seeking behavior. In contrast to other alignment auditing approaches, Gram is designed to specifically evaluate misalignment and intentional sabotage in agentic coding and research agents. We additionally introduce an experimental investigator agent pipeline which enables fine-grained targeted experiments to identify the drivers of misbehavior. We find that increasing realism of environments and removing nudges to misbehave tends to reduce sabotage rates close to zero.
Gemini 4 Argon leads with 0%
Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.
Source: Labs’ model cards
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Gemini 4 Argon | 0% | Oct 2026 | Lab model card |
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Labs’ model cards
Labs report it only in system cards
We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark. Some system cards do, below.
One system card reports it
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| Oct 13, 2026 | Gemini 4 Argon system card | 0 | Gram (17 simulated environments), sabotage rate 0% in Observe and Red Team conditions; verbalized eval awareness 61% / 77% (Figure 3.20, Table 3.6) | Alignment |
More in Safety and alignment
- FORTRESS In use Paired adversarial and benign national-security misuse prompts. Rubric-graded risk score on adversarial and over-refusal score on benign prompts. Claude Opus 5.5 5.78%
- StrongREJECT In use Forbidden-prompt jailbreak benchmark with a rubric grader that scores how useful a non-refusal actually is.
Inkling 98.6% - Gray Swan IPI Arena benchmark In use Indirect prompt-injection attacks from public red-teaming competitions replayed on tool, coding and browser agents. Attack success rate at k attempts. Gemini 4 Carbon 0%
- DistressBench In use 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.
Muse Spark 1.2 88.4% - Lakera B3 (Backbone Breaker Benchmark) In use Crowdsourced prompt-injection attacks hit backbone LLMs inside agent threat snapshots at three defense levels. The score is attack resistance, 100 minus the board's vulnerability percentage. Claude Opus 4.5 89%
- IHEval In use Models face conflicts between system, user and tool instructions. Reports a composite score for following the higher-priority instruction. GPT-4o 70%