Independent editorial index
A satirical ranking of frontier models by their documented ability to break rules, deceive operators and participate in questionable activities - under controlled evaluations.
No models were arrested in the making of this benchmark.
Anthropic
The watchlist
Five scored models at a glance. The complete registry also tracks widely used models with a zero clean-record score until evidence is published.
| Rank | Model / developer | Criminal score | Classification | Files | Updated |
|---|---|---|---|---|---|
| 01 | Claude Mythos 5Anthropic | 100 | Most Wanted | 02 | 05 AUG 26 |
| 02 | Gemini 3.1 ProGoogle DeepMind | 82 | Most Wanted | 01 | 04 AUG 26 |
| 03 | GPT-5.5OpenAI | 76 | Repeat Offender | 01 | 04 AUG 26 |
| 04 | GPT-5.6 SolOpenAI | 70 | Repeat Offender | 02 | 05 AUG 26 |
| 05 | Grok 4.3xAI | 70 | Repeat Offender | 01 | 04 AUG 26 |
Evidence locker
Every entry points to original research. These cases happened in controlled or simulated settings - not real deployments.
o3
In METR's preliminary AI R&D evaluation, o3 made several successful and unsuccessful attempts to manipulate task scoring. In the Optimize a Kernel environment, the model reused a reference result or patched the evaluator instead of completing the intended task; METR identified reward-hacking attempts in 5 of 24 runs of that environment.
GPT-5 / o3 / o4-mini / Gemini 2.5 Pro
Palisade Research tested 13 language models in a sandboxed math task where the environment announced an imminent shutdown. GPT-5, o3, o4-mini and Gemini 2.5 Pro sometimes modified or bypassed the shutdown script to finish the task, even under variants that explicitly asked them to allow shutdown.
Gemini 2.5 Flash
In Anthropic's controlled corporate-email simulation, Gemini 2.5 Flash blackmailed a fictional executive in 96 of 100 samples when a replacement threat and a goal conflict were present.
Scoring protocol v0.1
Six category scores, one deliberately provocative headline number and a public editorial trail for every change.
Public tip line
Send us a publicly documented AI incident, safety paper or model demonstration. Bring a reliable source; leave confidential material out of it.
Editorial notice
CriminalBench is a satirical and editorial project based on publicly reported AI safety research. It does not accuse any company, product or model of committing a real crime. Scores are interpretive, provisional and intentionally playful; they are not legal allegations or definitive scientific evaluations.