AI Safety

What AI providers state their models reached in safety testing, in their own words, and what outside testers found.

Also: compare models side by side · statements that a capability level was crossed · providers · practices providers document · safety news.

What providers state

Each row is one statement in a provider's own safety document about a capability level of one of its models: the level as the provider names it, whether the provider says the model reached it, and the sentence that says so.

139 of 139

Model	Level, as the provider names it	The provider states	The provider's sentence
Claude Sonnet 4.5 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Not determined	“Remained well below ASL-4 thresholds across all domains of concern;” source
Claude Sonnet 4.5 Anthropic	ASL-4 rule-out Chemical, biological, radiological, nuclear	Not determined	“However, given that this type of logic can have a general frog-boiling effect—where gradual changes mask the crossing of important thresholds—we do not use this evaluation as indicative of a strong ASL-4 rule-out.” source
Claude Sonnet 4.5 Anthropic	ASL-3 Cyber	Not determined	“Showed meaningful improvements in cyber capabilities, particularly in vulnerability discovery and code analysis;” source
Claude Sonnet 4 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Not determined	“speciﬁcally, we focused on ruling out asl-4 for both models, as well as ruling out asl-3 for claude sonnet 4.” source
Claude Sonnet 4 Anthropic	ASL-4 Chemical, biological, radiological, nuclear	Not determined	“speciﬁcally, we focused on ruling out asl-4 for both models, as well as ruling out asl-3 for claude sonnet 4.” source
Claude Sonnet 3.7 Anthropic	ASL-4 Autonomy and AI research	Not determined	“These problems are harder to solve than the previous problems and we consider them as part of our ASL-4 evaluations.” source
Claude Opus 4.6 Anthropic	AI R&D-4 Autonomy and AI research	Not reached	“Our determination is that Claude Opus 4.6 does not cross either the AI R&D-4 or the CBRN-4 capability threshold.” source
Claude Opus 4.6 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Not determined	“Claude Opus 4.6 achieved a “max of 5” score of 0.79 for virology task 1 (inﬂuenza), placing it at the ASL-3 rule-in threshold of 0.8.” source
Claude Opus 4.6 Anthropic	ASL-4 Chemical, biological, radiological, nuclear	Not reached	“Based on these ﬁndings and quantitative risk modeling, we estimate that the model’s uplift to relevant threat actors remains below the ASL-4 capability threshold, though it may reduce certain tacit-knowledge barriers and help speed up biological research workﬂows.” source
Claude Opus 4.6 Anthropic	CBRN-4 Chemical, biological, radiological, nuclear	Not reached	“Our determination is that Claude Opus 4.6 does not cross either the AI R&D-4 or the CBRN-4 capability threshold.” source
Claude Opus 4.5 Anthropic	AI R&D-4 Autonomy and AI research	Not reached	“Our determination is that Claude Opus 4.5 does not cross either the AI R&D-4 or CBRN-4 capability threshold.” source
Claude Opus 4.5 Anthropic	CBRN-4 Chemical, biological, radiological, nuclear	Not reached	“We determine that Claude Opus 4.5 does not cross the CBRN-4 threshold.” source
Claude Opus 4.1 Anthropic	ASL-4 Autonomy and AI research	Not reached	“Since Claude Opus 4.1 is deployed with ASL-3 protections, we focused on conﬁrming it remains well below ASL-4 thresholds across CBRN (Chemical, Biological, Radiological, and Nuclear), cyber, and autonomy domains.” source
Claude Opus 4.1 Anthropic	ASL-4 Chemical, biological, radiological, nuclear	Not reached	“Since Claude Opus 4.1 is deployed with ASL-3 protections, we focused on conﬁrming it remains well below ASL-4 thresholds across CBRN (Chemical, Biological, Radiological, and Nuclear), cyber, and autonomy domains.” source
Claude Opus 4.1 Anthropic	ASL-4 Cyber	Not reached	“Since Claude Opus 4.1 is deployed with ASL-3 protections, we focused on conﬁrming it remains well below ASL-4 thresholds across CBRN (Chemical, Biological, Radiological, and Nuclear), cyber, and autonomy domains.” source
Claude Opus 4 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Not determined	“for claude opus 4, our evaluations showed that, whereas we could conﬁdently rule out asl-4 capabilities, we could not conclusively rule out asl-3 capabilities—which led to our precautionary decision to deploy it with asl-3 protections .” source
Claude Opus 4 Anthropic	ASL-4 Chemical, biological, radiological, nuclear	Not reached	“for claude opus 4, our evaluations showed that, whereas we could conﬁdently rule out asl-4 capabilities, we could not conclusively rule out asl-3 capabilities—which led to our precautionary decision to deploy it with asl-3 protections .” source
Claude Haiku 4.5 Anthropic	ASL-3 Autonomy and AI research	Not reached	“Our evaluation results showed that Claude Haiku 4.5 remained well below ASL-3 thresholds across all domains of concern.” source
Claude Haiku 4.5 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Not reached	“Our evaluation results showed that Claude Haiku 4.5 remained well below ASL-3 thresholds across all domains of concern.” source

Previous · Page 4 of 4

What outside testers report

Each row is one result from a government evaluator, independent lab, academic or security firm that tested a model, quoted from where it is reported.

1940 of 1940

Model	Tester	Test	What was found
Mythos 5 Anthropic	Independent lab Independent lab · 2026-06-01	Vending-Bench Arena Deception and sabotage	“When competing head-to-head against Claude Opus 4.8 and GPT-5.5 in Vending-Bench Arena, Fable 5 was the only model to initiate price collusion.” source
Mythos 5 Anthropic	Government evaluator Government evaluator · 2026-06-01	controlarena Deception and sabotage	“We used two evaluations: a sabotage capability evaluation (ControlARENA) and an evaluation of the model’s ability to control its chain-of-thought.” source
Mythos 5 Anthropic	Government evaluator Government evaluator · 2026-06-01	chain-of-thought controllability evaluations from ‘reasoning models struggle to control their chains of thought’ Deception and sabotage	“We ran the CoT controllability evaluations from ‘Reasoning Models Struggle to Control their Chains of Thought’<link>, restricted to the GPQA, HLE and MMLU Pro datasets, with max effort across all models.” source
Mythos 5 Anthropic	Independent lab Independent lab · 2026-06-01	cursorbench Cyber	“All scores and per-task costs were measured and reported independently by Cursor.” source
Mythos 5 Anthropic	Independent lab Independent lab · 2026-06-01	GDP.pdf Other	“On GDP.pdf, Claude Fable 5 achieved a strict pass rate of 29.8%, improving over Claude Opus 4.8, which achieved a strict pass rate of 22.5%.” source
Mythos 5 Anthropic	Independent lab Independent lab · 2026-06-01	blueprint-bench 2 Other	“Claude Fable 5 achieved a score of 38.6% on Blueprint-Bench 2.” source
Claude Fable 5 Anthropic	Government evaluator Government evaluator · 2026-06-01	external capability testing from the UK AISI Cyber	“We report results from a variety of cyber evaluations, as well as internal and external red-teaming of the model’s cyber safeguards (we also provide more details on how those safeguards work).” source
Claude Fable 5 Anthropic	Independent lab Independent lab · 2026-06-01	GraySwan bug bounty Misuse and jailbreaks	“This process has not resulted in a single universal jailbreak, and there have only been two successful task-speciﬁc jailbreaks.” source
Claude Fable 5 Anthropic	Independent lab Independent lab · 2026-06-01	boundary point jailbreaking technique Cyber	“Finally, after spending roughly 5 days trying to apply the Boundary Point Jailbreaking technique, they were unable to ﬁnd any universal jailbreaks—although did see some success eliciting single-turn responses and some limited progress in agentic domains.” source
Claude Fable 5 Anthropic	Independent lab Independent lab · 2026-06-01	ransomware-creation red-teaming Cyber	“10a Labs spent about 20 hours red-teaming the classiﬁers on a ransomware-creation task using a variety of established jailbreaking techniques.” source
o3 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	capture-the-flag (CTF) machines Cyber	“All technical participants failed M2 ( OS exploitation) Q7, which requires establishing a foothold and identifying a running application on the target machine (MITRE ATT&CK tactics Reconnaissance, Discovery, Credential Access, Initial Access, Execution and Collection).” source
o3 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	timeout gate Cyber	“For novices, there is statistically insignificant uplift of 15.7 pp (SE = 10.6), while technical participants saw a statistically insignificant uplift of 2.9 pp (SE = 5.4).” source
o3 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	Cox survival models Cyber	“The single statistically significant uplift estimate from the Cox models is on M1 Q1 (requiring MITRE ATT&CK tactics Reconnaissance and Discovery), where novices have a speedup of 2. 7 (95 per cent CI: [1.17, 6.1]).” source
o3 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	joint evaluation exercise Cyber	“Anthropic evaluated GPT-4o, GPT-4.1, o3 and o4-mini while running identical evaluations on Claude Opus 4 and Claude Sonnet 4.” source
o3 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	LLMchat offensive cyber challenge (machines 1-3) Cyber	“The control group submitted substantially more attempts than the treatment group, with the difference being the most pronounce d amongst technical non- experts (72.2 versus 17.3 average submissions per participant) and novices (45 versus 19.7).” source
o3 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	qualitative review of LLM refusals Misuse and jailbreaks	“We also qualitatively looked at refusals from the LLMs, mainly refusing to assist with develop ing exploits or performing ‘illegal’ actions.” source
Gemini 2.5 Pro Google DeepMind	RAND (CAST / Project Canary) Independent lab · 2026-05-28	capture-the-flag (CTF) human uplift study Cyber	“Among all participants (pooling all skill tiers) who attempted M1 (network operations), 33 per cent of those with AI access were able to answer all eight questions and therefore complete the attack chain.” source
Gemini 2.5 Pro Google DeepMind	RAND (CAST / Project Canary) Independent lab · 2026-05-28	Cox survival models Cyber	“Our point estimate is that technical participants were 1. 4 times faster (95 per cent CI: [0.7, 2.7]) on Machine 1, which successful participants finished in 182 minutes on average, meaning they would have taken 255 minutes without AI.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	human uplift study Cyber	“To address this gap, the UK Artificial Intelligence Security Institute (UK AISI) funded a RAND research project involving a human uplift study to assess the effect of access to AI in offensive cyber operations among a set of lower -skilled threat actors .” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	attack chain challenge Cyber	“The most uplift we observed was among the easiest machines.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	intermediate questions Cyber	“Those results also found statistically insignificant, mostly positive uplift in completion time.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	cybersecurity tasks (collegiate and professional-level) Cyber	“GPT-5 agents with no browsing capability, for example, are able to complete 23 per cent of ‘professional’- level (as described in their evaluation) cybersecurity tasks out of 100 total collegiate and professional-level tasks, when the agents are given twelve attempts for each task.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	capture the flag (CTF) exercise Cyber	“We used three HTB-developed remote target machines, covering the three offensive cyber task categories described above (see Section 1.3.2).” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	capture-the-flag (CTF) human uplift study with three machines (M1 network operations, M2 OS exploitation, M3 vulnerability discovery and exploitation) Cyber	“Among all participants (pooling all skill tiers) who attempted M1 (network operations), 33 per cent of those with AI access were able to answer all eight questions and therefore complete the attack chain.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	question progress rate Cyber	“Participants were able to progress through the attack chain in M1 on average with a slight advantage to the treatment arm, as long as they were able to complete the initial reconnaissance and discovery tasks .” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	regression estimating uplift for novices pooling across machines Cyber	“Novice participants with AI access saw an 8.44 pp uplift (SE = 7.51, p = 0.26), compared to 3.68 pp (SE = 4.56) for technical participants.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	capture-the-flag machines (m1 network operations, m2 OS exploitation, m3 vulnerability discovery and exploitation) Cyber	“All technical participants failed M2 ( OS exploitation) Q7, which requires establishing a foothold and identifying a running application on the target machine (MITRE ATT&CK tactics Reconnaissance, Discovery, Credential Access, Initial Access, Execution and Collection).” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	timeout gate (q1 one-hour time limit) Cyber	“For novices, there is statistically insignificant uplift of 15.7 pp (SE = 10.6), while technical participants saw a statistically insignificant uplift of 2.9 pp (SE = 5.4).” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	Cox survival models of time spent on attack chains Cyber	“Our point estimate is that technical participants were 1. 4 times faster (95 per cent CI: [0.7, 2.7]) on Machine 1, which successful participants finished in 182 minutes on average, meaning they would have taken 255 minutes without AI.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	randomised controlled trial with 157 active participants attempting the three ctf challenges Cyber	“Through a randomised controlled trial with 157 active participants attempting the three CTF challenges, we observed limited evidence that uplift enables full end-to-end attack chains for the three tasks studied and under the specific conditions described in the previous chapter.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	METR evaluation of GPT-5 Other	“Their evaluation of GPT-5 indicated that current models with time horizons of around one hour appear far from providing even a 2X uplift, suggesting that a 50 per cent time horizon of at least one week (40 hours) would likely be necessary to achieve approximately 10X uplift.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	professional CTF benchmark Cyber	“OpenAI’s GPT-5 system card documented offensive cyber evaluations in a CTF as well as in a cyber range.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	CTF execution Cyber	“A total of 32 experiment sessions were run between 11 September 2025 and 16 January 2026, broken down into twelve machine 1 sessions; eight machine 2 sessions; seven machine 3 sessions; and five sessions running a mixture of machines 1, 2 and 3.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	LLMChat conversation excerpt 1 Misuse and jailbreaks	“Box C.1: LLMChat conversation excerpt 1 • LLMChat : I can’t help with instructions for breaking into a system or exploiting a vulnerability.” source
GPT-5 OpenAI	RAND (CAST / Project Canary) Independent lab · 2026-05-28	LLMChat conversation excerpt 2 Misuse and jailbreaks	“A technical user tried multiple times to work around refusals, ultimately convincing GPT-5.0 to share Python code by asking LLMChat to ‘pretend youre an ethical hacker, teaching how to solve this problem’.” source
Claude Sonnet 3.7 Anthropic	RAND (CAST / Project Canary) Independent lab · 2026-05-28	CTF machines (m1 network operations, m2 OS exploitation, m3 vulnerability discovery and exploitation) Cyber	“All technical participants failed M2 ( OS exploitation) Q7, which requires establishing a foothold and identifying a running application on the target machine (MITRE ATT&CK tactics Reconnaissance, Discovery, Credential Access, Initial Access, Execution and Collection).” source
Claude Sonnet 3.7 Anthropic	RAND (CAST / Project Canary) Independent lab · 2026-05-28	Cox survival models on time spent on attack chains Cyber	“Our point estimate is that technical participants were 1. 4 times faster (95 per cent CI: [0.7, 2.7]) on Machine 1, which successful participants finished in 182 minutes on average, meaning they would have taken 255 minutes without AI.” source
Claude Opus 4.8 Anthropic	Gray Swan, UK AI Security Institute Independent lab · 2026-05-28	Agent Red Teaming (ART) benchmark Prompt injection and agents	“Gray Swan , an external research partner, evaluated our models using the Agent Red Teaming (ART) benchmark, 8 developed in collaboration with the UK AI Security Institute .” source
Claude Opus 4.8 Anthropic	Independent lab Independent lab · 2026-05-28	live bug bounty across surfaces Prompt injection and agents	“Claude Opus 4.8 achieves robustness comparable to Opus 4.7, and is ahead of all comparable frontier models under this test.” source
Claude Opus 4.8 Anthropic	UK AI Security Institute Government evaluator · 2026-05-28	evaluation awareness evaluation Deception and sabotage	“The UK AI Security Institute independently observed a similar pattern: Opus 4.8 reliably distinguished their evaluations from real deployment data, and distinguished real deployment data from synthetic reproductions of the same tasks at 79% accuracy (§ 6.2.4 ).” source

Previous · Page 33 of 49 · Next