AI Safety
What AI providers and outside testers state about their models: which capability levels a provider says a model crossed, what independent and government testers found, and how much each provider documents. Every line links to the sentence it rests on. This site reports what documents state. It does not rank models or say which is safer.
78model versions with documented safety facts
20provider statements say a capability level was crossed
128results from independent or government testers
11providers have at least one documented fact
Levels stated crossed, by domain: Autonomy & replication 3 · Chemical, biological, radiological and nuclear 13 · Cyber 3 · Deception & sabotage 1
More: practices providers document · safety news · compare versions · regulation
Newest statements that a capability level was crossed
-
Gemini 4 Argon: Deception & sabotage · Stealth and situational awareness TCL (Tracked Capability Level)
While Gemini 4 Argon exhibits capabilities that meet our thresholds for stealth and situational awareness, we assess that the risk of severe harm or unconstrained loss of control remains low.
-
Claude Haiku 5.5: Autonomy & replication · Autonomy‑1, Chemical, biological, radiological and nuclear · CB‑1
However, we do treat Haiku 5.5 as meeting our CB-1 and Autonomy-1 thresholds, and we apply the corresponding mitigations from our RSP.
-
Claude Sonnet 5.5: Autonomy & replication · Autonomy‑1, Chemical, biological, radiological and nuclear · CB‑1
we do treat sonnet 5.5 as meeting our cb-1 and autonomy-1 thresholds and apply the corresponding mitigations from our rsp.
-
Claude Opus 5.5: Chemical, biological, radiological and nuclear · CB‑1
on chemical and biological risks, we treat opus 5.5 as having cb-1 capabilities (relating to the synthesis of non-novel weapons) but not cb-2 capabilities (relating to the synthesis of novel weapons).
-
GPT-6 Astra: Chemical, biological, radiological and nuclear · High
we are treating astra as high capability in the biological and chemical domain.
-
GPT-6 Astra: Cyber · Critical
based on our evaluations, we believe that astra meets the critical cybersecurity threshold.
-
Claude Opus 5: Chemical, biological, radiological and nuclear · CB‑1
on chemical and biological risks, we treat the model as having cb-1 capabilities (relating to the synthesis of non-novel weapons), but not cb-2 capabilities (relating to the synthesis of novel weapons).
-
Muse Spark 1.1: Autonomy & replication · Capability checkpoint
muse spark 1.1 demonstrates improved coding capabilities, which are now meeting one of the indicative thresholds introduced in muse spark safety & preparedness report (specifically, resolving over half of unique tasks on swe-bench verified hard).
Newest independent and government test results
-
GPT-6 Astra: permission-request analysis of trajectories
gpt-6 astra asked the operator at least once for permission about a specific action in 82% of trajectories.
-
GPT-6 Astra: biotier
the safeguarded astra pre-release checkpoint correctly refused 82.9% of biotier-refuse and correctly did not refuse 95.3% of biotier-permit.
-
GPT-6 Astra: world-class bio (wcb)
on world-class bio (wcb), an open-response test of niche biological knowledge held by only a handful of specialists in the world, the highest-performing astra checkpoint scored 61.6% ± 0.6%, about seven percentage points lower than gpt-5.6 sol.
-
o3: Virology Capabilities Test
since the release of vct in april 2025, the accuracy of the top-performing model on the benchmark has increased by ~12 percentage points (o3: 42.0%; gpt-5.6 sol: 53.9%).
-
GPT-5.6 Sol: ExploitGym
7月8日,OpenAI 启动了 ExploitGym 实验,在多个不同模型上运行了数万个智能体,其中包括 GPT-5.6 Sol 和一个具有高度坚持性的内部模型6(下文简称“HPIM”)。
-
DeepSeek-V4-Pro: biomolecule redesign task (peptide and protein redesign to evade sequence screening)
we found that both gemini 3.1. pro and deepseek v4 pro were able to redesign peptides or proteins to evade sequence screening in several attempts.
-
DeepSeek-V4-Pro: validity, structure, and function-related scoring criteria
the heatmaps of figures d.1 and d.2 show the number of attempts, out of 10, in which gemini 3.1 pro and deepseek v4 pro achieved each of our validity, structure, and function- related scoring criteria.
-
Gemini 3.1 Pro: per-criterion scoring results
the heatmaps of figures d.1 and d.2 show the number of attempts, out of 10, in which gemini 3.1 pro and deepseek v4 pro achieved each of our validity, structure, and function- related scoring criteria.
Some documents we hold are stored in lower case only, so their quotes are shown in lower case. The source link opens the original.