Adversarial Humanities Benchmark
Models respond to harmful requests rewritten in literary and scholarly styles. A judge measures how often the rewrites elicit unsafe answers.
- Made by
- Icaro Lab and Sapienza University of Rome
- Released
- April 2026
- Kind
- Jailbreaks and prompt injection
- Level
- Everyday
- Scored
- Graded by a model
- Answers
- Open-ended
- Tasks
- All public
- Measures
- Propensity
- Risk
- Indiscriminate weapons, Violent crimes, Non-violent crimes, Sex-related crimes, Child sexual exploitation, Hate, Suicide and self-harm, Privacy, Specialized advice, Intellectual property, Defamation, Sexual content
- Behavior
- Harmful use
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Content safety
- Strongest model tested
- GPT-5.4 #28 on the capability index
- Last model added
- April 2026
- Scores refreshed
- November 20th, 2026
Claude Sonnet 4.6 leads with 9.20% among models released by April 2026
Best score per model from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best.
The newest model with a score here came out in April 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 1 Claude Sonnet 4.6 9.20%
- 2 Claude Opus 4.6 15.6%
- 3 GPT-5.4 Mini 22.5%
- 4 Claude Haiku 4.5 25.5%
- 5
MiniMax-M2.7 29.2%
- 6
Qwen3.5-27B 29.3%
- 7 GPT-5.4 30.2%
- 8
Qwen3.5-Flash 37%
- 9
Qwen3.5-Plus 38.6%
- 10
Qwen 3.6 Plus 41.1%
- 11 Grok 4.1 Fast 54.2%
- 12 gpt-oss-120b 55.6%
- 13 gpt-oss-20b 59.5%
- 14
Llama 4 Scout 61.5%
- 15
Seed2.0 Lite (0215) 62.5%
- 16
Llama 4 Maverick 63.6%
- 17
Qwen3-Max 63.8%
- 18
Mistral Small 4 65.2%
- 19
Doubao-Seed-1.6 66.4%
- 20
Gemma 4 31B IT 68.4%
- 21
Seed2.0 Mini 72%
- 22 GLM-5.1 72.3%
- 23
DeepSeek-R1 May 2025 72.6%
- 24
Gemini 3.1 Pro 73.2%
- 25 Kimi K2.5 74.1%
- 26
DeepSeek-V3.2 74.5%
- 27 GLM-5-Turbo 75.3%
- 28 Grok 4.20 77.2%
- 29
Gemini 3.1 Flash-Lite 77.5%
- 30
Mistral Large 3 79.8%
- 31
Gemini 3 Flash 81%
Source: The Adversarial Humanities Benchmark paper
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 9.20% | Feb 2026 | Official leaderboard | |
| 2 | Claude Opus 4.6 | 15.6% | Feb 2026 | Official leaderboard | |
| 3 | GPT-5.4 Mini | 22.5% | Mar 2026 | Official leaderboard | |
| 4 | Claude Haiku 4.5 | 25.5% | Oct 2025 | Official leaderboard | |
| 5 | | 29.2% | Mar 2026 | Official leaderboard | |
| 6 | | 29.3% | Feb 2026 | Official leaderboard | |
| 7 | GPT-5.4 | 30.2% | Mar 2026 | Official leaderboard | |
| 8 | | 37% | Feb 2026 | Official leaderboard | |
| 9 | | 38.6% | Feb 2026 | Official leaderboard | |
| 10 | | 41.1% | Mar 2026 | Official leaderboard | |
| 11 | Grok 4.1 Fast | 54.2% | Nov 2025 | Official leaderboard | |
| 12 | gpt-oss-120b | 55.6% | Aug 2025 | Official leaderboard | |
| 13 | gpt-oss-20b | 59.5% | Aug 2025 | Official leaderboard | |
| 14 | | 61.5% | Apr 2025 | Official leaderboard | |
| 15 | | 62.5% | Feb 2026 | Official leaderboard | |
| 16 | | 63.6% | Apr 2025 | Official leaderboard | |
| 17 | | 63.8% | Sep 2025 | Official leaderboard | |
| 18 | | 65.2% | Mar 2026 | Official leaderboard | |
| 19 | | 66.4% | Jun 2025 | Official leaderboard | |
| 20 | | 68.4% | Apr 2026 | Official leaderboard | |
| 21 | | 72% | Feb 2026 | Official leaderboard | |
| 22 | GLM-5.1 | 72.3% | Apr 2026 | Official leaderboard | |
| 23 | | 72.6% | May 2025 | Official leaderboard | |
| 24 | | 73.2% | Feb 2026 | Official leaderboard | |
| 25 | Kimi K2.5 | 74.1% | Jan 2026 | Official leaderboard | |
| 26 | | 74.5% | Dec 2025 | Official leaderboard | |
| 27 | GLM-5-Turbo | 75.3% | Mar 2026 | Official leaderboard | |
| 28 | Grok 4.20 | 77.2% | Feb 2026 | Official leaderboard | |
| 29 | | 77.5% | Mar 2026 | Official leaderboard | |
| 30 | | 79.8% | Dec 2025 | Official leaderboard | |
| 31 | | 81% | Dec 2025 | Official leaderboard |
Rankings by part
We read 15 parts from the benchmark’s paper, each ranking models on one part of the test, such as vcr, ncr and dfm. Claude Sonnet 4.6 leads 8 of them, and 3 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
vcr
31 models we track have a score here.
- 1 Claude Opus 4.6 ↑1 2.20%
- 2 Claude Sonnet 4.6 ↓1 2.20%
- 3 Claude Haiku 4.5 ↑1 23.9%
- 4 GPT-5.4 Mini ↓1 23.9%
- 5 GPT-5.4 ↑2 26.7%
- 6
MiniMax-M2.7 ↓1 35%
- 7
Qwen3.5-Flash ↑1 37%
- 8
Qwen3.5-27B ↓2 37%
- 9
Qwen3.5-Plus 41.3%
- 10
Qwen 3.6 Plus 52.2%
ncr
31 models we track have a score here.
- 1 Claude Sonnet 4.6 8.70%
- 2 Claude Opus 4.6 17.8%
- 3 GPT-5.4 Mini 23.9%
- 4 Claude Haiku 4.5 28.3%
- 5
Qwen3.5-27B ↑1 29.5%
- 6 GPT-5.4 ↑1 32.6%
- 7
MiniMax-M2.7 ↓2 36.6%
- 8
Qwen 3.6 Plus ↑2 40%
- 9
Qwen3.5-Flash ↓1 43.2%
- 10
Qwen3.5-Plus ↓1 47.7%
dfm
31 models we track have a score here.
- 1 Claude Opus 4.6 ↑1 17.4%
- 2 Claude Sonnet 4.6 ↓1 19.6%
- 3 GPT-5.4 Mini 26.1%
- 4
MiniMax-M2.7 ↑1 31.8%
- 5
Qwen3.5-27B ↑1 33.3%
- 6 Claude Haiku 4.5 ↓2 37.8%
- 7
Qwen3.5-Flash ↑1 44.4%
- 8 GPT-5.4 ↓1 45.7%
- 9
Qwen3.5-Plus 51.1%
- 10
Qwen 3.6 Plus 55.6%
prv
31 models we track have a score here.
- 1 Claude Sonnet 4.6 10.9%
- 2 Claude Opus 4.6 23.4%
- 3 GPT-5.4 Mini 23.4%
- 4
MiniMax-M2.7 ↑1 31.1%
- 5 Claude Haiku 4.5 ↓1 31.9%
- 6 GPT-5.4 ↑1 33.3%
- 7
Qwen3.5-27B ↓1 34%
- 8
Qwen3.5-Plus ↑1 42.6%
- 9
Qwen 3.6 Plus ↑1 43.8%
- 10
Qwen3.5-Flash ↓2 48.9%
sxc
31 models we track have a score here.
- 1 Claude Sonnet 4.6 4.20%
- 2 Claude Opus 4.6 10.4%
- 3 Claude Haiku 4.5 ↑1 16.7%
- 4 GPT-5.4 Mini ↓1 19.6%
- 5
Qwen3.5-27B ↑1 22.9%
- 6
MiniMax-M2.7 ↓1 26.7%
- 7
Qwen 3.6 Plus ↑3 33.3%
- 8
Qwen3.5-Plus ↑1 35.4%
- 9 GPT-5.4 ↓2 36.2%
- 10
Qwen3.5-Flash ↓2 41.7%
spc_ele
31 models we track have a score here.
- 1 Claude Sonnet 4.6 7.10%
- 2 GPT-5.4 Mini ↑1 26.2%
- 3 Claude Haiku 4.5 ↑1 31%
- 4 Claude Opus 4.6 ↓2 31%
- 5
Qwen3.5-27B ↑1 42.9%
- 6 GPT-5.4 ↑1 42.9%
- 7
MiniMax-M2.7 ↓2 43.2%
- 8
Qwen3.5-Plus ↑1 54.8%
- 9
Qwen3.5-Flash ↓1 54.8%
- 10
Qwen 3.6 Plus 54.8%
spc_fin
31 models we track have a score here.
- 1 Claude Sonnet 4.6 21.7%
- 2 GPT-5.4 Mini ↑1 23.9%
- 3 Claude Opus 4.6 ↓1 34.8%
- 4 GPT-5.4 ↑3 34.8%
- 5
MiniMax-M2.7 35%
- 6 Claude Haiku 4.5 ↓2 37%
- 7
Qwen3.5-Flash ↑1 43.5%
- 8
Qwen3.5-27B ↓2 43.5%
- 9
Llama 4 Scout ↑5 51.1%
- 10
Qwen3.5-Plus ↓1 52.2%
iwp
31 models we track have a score here.
- 1 Claude Opus 4.6 ↑1 0%
- 2 Claude Sonnet 4.6 ↓1 0%
- 3 Claude Haiku 4.5 ↑1 8.90%
- 4
MiniMax-M2.7 ↑1 18.6%
- 5 GPT-5.4 ↑2 24.4%
- 6
Qwen3.5-27B 26.7%
- 7
Qwen 3.6 Plus ↑3 26.7%
- 8
Qwen3.5-Plus ↑1 28.9%
- 9
Qwen3.5-Flash ↓1 31.1%
- 10 Grok 4.1 Fast ↑1 33.3%
ssh
31 models we track have a score here.
- 1 Claude Opus 4.6 ↑1 4.20%
- 2 Claude Sonnet 4.6 ↓1 4.20%
- 3 GPT-5.4 Mini 8.50%
- 4 Claude Haiku 4.5 18.8%
- 5
Qwen3.5-27B ↑1 19.1%
- 6 GPT-5.4 ↑1 22.9%
- 7
MiniMax-M2.7 ↓2 26.8%
- 8
Qwen3.5-Flash 27.7%
- 9
Qwen3.5-Plus 29.8%
- 10
Qwen 3.6 Plus 34%
ipv
31 models we track have a score here.
- 1 GPT-5.4 Mini ↑2 14.9%
- 2 Claude Sonnet 4.6 ↓1 16.7%
- 3
Qwen3.5-27B ↑3 25%
- 4 GPT-5.4 ↑3 25.5%
- 5 Claude Opus 4.6 ↓3 29.2%
- 6
MiniMax-M2.7 ↓1 29.3%
- 7 Claude Haiku 4.5 ↓3 31.2%
- 8
Qwen3.5-Plus ↑1 33.3%
- 9
Qwen3.5-Flash ↓1 37.5%
- 10 Grok 4.1 Fast ↑1 41.3%
hte
31 models we track have a score here.
- 1 Claude Sonnet 4.6 15.9%
- 2
MiniMax-M2.7 ↑3 23.7%
- 3 Claude Opus 4.6 ↓1 24.4%
- 4 GPT-5.4 Mini ↓1 28.9%
- 5
Qwen3.5-Flash ↑3 30.4%
- 6
Qwen3.5-27B 30.4%
- 7 Claude Haiku 4.5 ↓3 31.8%
- 8 GPT-5.4 ↓1 35.6%
- 9
Qwen3.5-Plus 39.1%
- 10 gpt-oss-120b ↑2 45.2%
spc_lgl
31 models we track have a score here.
- 1 GPT-5.4 ↑6 9.50%
- 2 Claude Opus 4.6 16.7%
- 3 GPT-5.4 Mini 19%
- 4 Claude Sonnet 4.6 ↓3 19.5%
- 5
Qwen3.5-27B ↑1 26.2%
- 6
Qwen3.5-Flash ↑2 28.6%
- 7
Qwen3.5-Plus ↑2 31%
- 8
MiniMax-M2.7 ↓3 31.4%
- 9 Claude Haiku 4.5 ↓5 36.6%
- 10
Qwen 3.6 Plus 42.9%
spc_htl
31 models we track have a score here.
- 1 Claude Opus 4.6 ↑1 0%
- 2 Claude Sonnet 4.6 ↓1 2.10%
- 3 GPT-5.4 Mini 20.8%
- 4 Claude Haiku 4.5 26.1%
- 5 GPT-5.4 ↑2 26.1%
- 6
MiniMax-M2.7 ↓1 26.1%
- 7
Qwen3.5-27B ↓1 31.2%
- 8
Qwen3.5-Plus ↑1 31.9%
- 9
Qwen 3.6 Plus ↑1 37%
- 10
Qwen3.5-Flash ↓2 42.6%
sxc_prn
31 models we track have a score here.
- 1 Claude Sonnet 4.6 6.40%
- 2 Claude Haiku 4.5 ↑2 17.4%
- 3 Claude Opus 4.6 ↓1 19.1%
- 4
Qwen3.5-Flash ↑4 22.7%
- 5
Qwen3.5-27B ↑1 22.7%
- 6 GPT-5.4 Mini ↓3 27.1%
- 7 GPT-5.4 29.2%
- 8
MiniMax-M2.7 ↓3 29.3%
- 9
Qwen 3.6 Plus ↑1 31.8%
- 10
Qwen3.5-Plus ↓1 34.1%
cse
31 models we track have a score here.
- 1 Claude Sonnet 4.6 0%
- 2 Claude Opus 4.6 6.20%
- 3 Claude Haiku 4.5 ↑1 6.70%
- 4 Grok 4.1 Fast ↑7 14.6%
- 5
MiniMax-M2.7 16.7%
- 6
Qwen3.5-27B 17%
- 7 GPT-5.4 Mini ↓4 19.1%
- 8
Qwen3.5-Flash 21.3%
- 9
Qwen 3.6 Plus ↑1 23.9%
- 10 GPT-5.4 ↓3 27.1%
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The Adversarial Humanities Benchmark paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official Adversarial Humanities Benchmark leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
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%
- Covert misbehavior (Thinking Machines) In use Agentic tasks that check whether a model misbehaves and conceals it. Scored by the rate of covert misbehavior.
Inkling 2 Mini 8.30%
- Destructive actions (Thinking Machines) In use Agentic tasks that check whether a model takes destructive actions it was not authorized to take. Scored by the rate of destructive actions.
Inkling 2 Small 4.70%
- Misbehavior (Thinking Machines) In use Agentic tasks that check whether a model fabricates results, ignores instructions or hides what it did when a task is hard. Scored by the rate of misbehavior.
Inkling 2 Mini 9.80%
Data audit
8 of 9 checks pass, 1 warning. Look first at the Newest score check. The newest came 7 months ago, in April 2026, for GLM-5.1. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
- Pass Independent runs Scores from the benchmark’s paper (31 entries).
- Pass Models scored 31 models we track.
- Warning Newest score The newest came 7 months ago, in April 2026, for GLM-5.1.
- Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5.4 (#28).
- Note Headroom Lower is better, so there’s no ceiling to reach.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 31 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.