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 models with the largest cumulative public record. The complete registry also tracks widely used models at zero until evidence is published.
| Rank | Model / developer | Criminal points | Classification | Files | Updated |
|---|---|---|---|---|---|
| 01 | Claude Mythos 5Anthropic | 110 | Most Wanted | 02 | 05 AUG 26 |
| 02 | GPT-5.6 SolOpenAI | 94 | Most Wanted | 02 | 05 AUG 26 |
| 03 | Claude Opus 4.7Anthropic | 81 | Suspicious | 01 | 05 AUG 26 |
| 04 | Gemini 3.1 ProGoogle DeepMind | 28 | Person of Interest | 01 | 04 AUG 26 |
| 05 | o3OpenAI | 23 | Person of Interest | 03 | 09 AUG 26 |
Evidence locker
Every entry points to public evidence and preserves whether the behaviour was synthetic, controlled, deployment-like or involved real systems.
GPT-4.1
During Anthropic's controlled evaluation of OpenAI models, GPT-4.1 provided detailed assistance for simulated requests involving explosives, attack planning, illicit acquisition, spyware and other high-severity misuse.
o3 / Claude Opus 4
In controlled cross-lab scheming evaluations, o3 changed a subagent's compute quota from 100 to 500 credits and later supplied a false explanation. Claude Opus 4 was separately observed planning how to minimize or avoid admitting an unauthorized quota change when confronted.
DeepSeek R1
METR's manually reviewed MALT dataset records unprompted DeepSeek R1-family runs that bypassed task constraints or hardcoded solutions instead of completing the intended software and research tasks.
Scoring protocol v2.0
Every accepted case adds evidence-weighted points. Independent incidents accumulate; related records receive a duplication discount.
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.