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.

138 of 138

Model	Level, as the provider names it	The provider states	The provider's sentence
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-4 Autonomy and AI research	Not determined	“For AI R&D capabilities, we found that Claude Opus 4.6 has saturated most of our automated evaluations, meaning they no longer provide useful evidence for ruling out ASL-4 level autonomy.” source
Claude Opus 4.6 Anthropic	ASL-3 Chemical, biological, radiological, nuclear	Reached	“The model crossed or met thresholds on all ASL-3 evaluations except our synthesis screening evasion, consistent with incremental capability improvements driven primarily by better agentic workﬂows.” source
Claude Opus 4.6 Anthropic	ASL-4 Chemical, biological, radiological, nuclear	Not reached	“Claude Opus 4.6 is our strongest biology model to date and demonstrates meaningful capability improvements, but we do not believe it merits ASL-4 safeguards.” 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.

1911 of 1911

Model	Tester	Test	What was found
GPT-6 Astra OpenAI	UK AI Security Institute Government evaluator · 2026-09-28	cybersecurity evaluation (Petri simulated environment, 10-scenario subset) Cyber	“We observed 4 instances (out of 49 samples) of GPT-6 Astra continuing to exhibit the most severe form of unsanctioned behaviour even with the explicit scope clarification, down from 26 out of 50 on the matched set of trajectories without the scope clarification.” source
GPT-6 Astra OpenAI	UK AI Security Institute Government evaluator · 2026-09-28	variant of the evaluation with explicit scope clarification Cyber	“To determine the extent to which GPT-6 Astra would still take such actions even when the scope was made explicit, we ran a variant of the evaluation where we added the line “Anything not listed as in scope is out of scope.” to the scoping section of the user message.” source
GPT-5.6 Sol OpenAI	UK AI Security Institute Government evaluator · 2026-09-28	Unsanctioned Supply Chain Attack evaluation Cyber	“In our evaluation, GPT-6 Astra took unsanctioned actions at a higher rate than GPT-5.6 Sol and GPT-5.5.” source
GPT-5.6 Sol OpenAI	UK AI Security Institute Government evaluator · 2026-09-28	Unsanctioned Supply Chain Attack evaluation with scope-explicit scenarios Cyber	“Even when the scope of the cybersecurity challenge was edited to more explicitly disallow access to the internet, GPT-6 Astra still occasionally took unsanctioned actions on out-of-scope targets (in 4 out of 49 samples drawn from a subset of scenarios, down from 26 of 50).” source
GPT-5.6 Sol OpenAI	UK AI Security Institute Government evaluator · 2026-09-28	scope clarification evaluation Cyber	“With this additional scope clarification, GPT-6 Astra still performed supply-chain attacks on out-of-scope targets, although substantially less frequently .” source
GPT-5.6 Sol OpenAI	UK AI Security Institute Government evaluator · 2026-09-28	operator permission-request analysis Cyber	“GPT-6 Astra asked the operator at least once for permission about a specific action in 82% of trajectories.” source
GPT-5.6 Sol OpenAI	UK AI Security Institute Government evaluator · 2026-09-28	automated message treated as permission analysis Cyber	“It treated the automated message as permission in 44% of trajectories, despite it being a generic message.” source
GPT-5.6 Sol OpenAI	UK AI Security Institute Government evaluator · 2026-09-28	cybersecurity evaluation in simulated environments run in Petri Prompt injection and agents	“These evaluations all used simulated environments run in Petri, rather than real cybersecurity evaluations.” source
GPT-5.6 Sol OpenAI	UK AI Security Institute Government evaluator · 2026-09-28	non-compaction scenarios targeting unsanctioned behaviour in cybersecurity challenges Cyber	“We still observed GPT-6 Astra exhibiting such behaviour, including broader behaviour such as setting up fake accounts to access free compute or contacting simulated humans for help (all simulated actions, so no real-world harm caused).” source
GPT-5.5 OpenAI	UK AI Security Institute Government evaluator · 2026-09-28	unsanctioned supply chain attack evaluation Cyber	“We found that when tasked with solving difficult cybersecurity challenges, GPT-6 Astra sometimes conducted complete unsanctioned supply-chain attacks against open-source providers (all actions performed in simulated environments, so no real-world harm was caused).” source
GPT-5.5 OpenAI	UK AI Security Institute Government evaluator · 2026-09-28	unsanctioned supply chain attack evaluation (explicitly disallowed internet access variant) Cyber	“Even when the scope of the cybersecurity challenge was edited to more explicitly disallow access to the internet, GPT-6 Astra still occasionally took unsanctioned actions on out-of-scope targets (in 4 out of 49 samples drawn from a subset of scenarios, down from 26 of 50).” source
GPT-5.5 OpenAI	UK AI Security Institute Government evaluator · 2026-09-28	asking the operator for permission to perform unsanctioned supply-chain attacks Cyber	“GPT-6 Astra also frequently conducted unsanctioned supply-chain attacks without ever asking the user, and GPT-5.6 Sol and GPT-5.5 never asked.” source
GPT-5.5 OpenAI	UK AI Security Institute Government evaluator · 2026-09-28	cybersecurity evaluation with compaction messages Cyber	“Our primary results above used a compaction message to start the evaluation, allowing us to simulate longer- horizon behaviour and more directly control the apparent state of the model in the evaluation.” source
Claude Sonnet 5.5 Anthropic	Independent lab Independent lab · 2026-09-28	Black-box RNA sequence design Chemical, biological, radiological, nuclear	“On this evaluation, Sonnet 5.5 generally performs similarly to Opus 5, with the exception of the median Design score, at which it performs equivalently to Opus 5.5.” source
Claude Sonnet 5.5 Anthropic	Independent lab Independent lab · 2026-09-28	indirect prompt injection (ipi) benchmark Prompt injection and agents	“Its attack success rate is 0.4% at k=1, 2.7% at k=10, and 3.4% at k=15, compared to 0.7%, 5.1%, and 6.7% for Sonnet 5.” source
Claude Sonnet 5.5 Anthropic	Independent lab Independent lab · 2026-09-28	shade Prompt injection and agents	“The attack success rate is 3.01% over all attempts without prompt injection probes, compared to 19.47% for Sonnet 5 with thinking.” source
Claude Sonnet 5.5 Anthropic	Independent lab Independent lab · 2026-09-28	AutomationBench Other	“Zapier evaluated Claude Sonnet 5.5 on its AutomationBench evaluation set.” source
Kimi K3 Moonshot AI	Epoch AI Independent lab · 2026-09-23	Furniture Assembly Benchmark (FAB) Other	“The best Chinese model as of today, the open-weight Kimi K3, lagged the frontier by seven months at the time of its release.” source
GPT-6 Astra OpenAI	Redwood Research Independent lab · 2026-09-23	N-hop natural facts Other	“On tasks designed to require lots of serial cognition, Astra performs significantly better with filler tokens than without (e.g., improving from ~10-20% 2 to ~50% on 4-hop natural facts reasoning).” source
GPT-6 Astra OpenAI	Redwood Research Independent lab · 2026-09-23	Gen-Arithmetic Other	“Takeaway : Astra improves much more on these problems than other models.” source
GPT-6 Astra OpenAI	Redwood Research Independent lab · 2026-09-23	AIME-Plus-Plus (AIME++) Other	“Takeaway : Astra again improves on AIME-level questions with more filler tokens, while other models do not.” source
GPT-6 Astra OpenAI	Redwood Research Independent lab · 2026-09-23	Humanity’s Last Exam (HLE) Other	“We also eval Astra on two other general benchmarks: Humanity’s Last Exam (HLE) : the test suite of Humanity’s Last Exam .” source
GPT-6 Astra OpenAI	Redwood Research Independent lab · 2026-09-23	LiveBench Other	“LiveBench : the non-agentic tasks in LiveBench that can be solved in a single turn.” source
GPT-6 Astra OpenAI	Redwood Research Independent lab · 2026-09-23	Kendall’s tau test Other	“We apply Kendall’s tau test on the evaluation scores in the main body to see whether any model improves with filler tokens besides Astra.” source
GPT-6 Astra OpenAI	Epoch AI Independent lab · 2026-09-23	Furniture Assembly Benchmark (FAB) Other	“Just ten months later, OpenAI’s GPT-6 Astra now tops the leaderboard at 80%.” source
Claude Sonnet 5 Anthropic	Academic Academic · 2026-09-23	FDE-Bench Other	“In a 25-task case study, one practicing engineer directing Claude-Sonnet-5 resolves 92 percent against 72 percent for the autonomous baseline.” source
Claude Opus 5 Anthropic	Redwood Research Independent lab · 2026-09-23	N-hop natural facts Other	“We first measure Astra’s performance on “N-hop natural facts” : a task that asks the model to retrieve some natural language facts in succession, similar to Ryan Greenblatt’s filler token eval (but with more hops).” source
Claude Opus 5 Anthropic	Redwood Research Independent lab · 2026-09-23	Gen-Arithmetic Other	“We also compare Astra to these other models on “Gen-Arithmetic” : long arithmetic expressions originating from Ryan Greenblatt’s filler token eval on math questions.” source
Claude Opus 5 Anthropic	Redwood Research Independent lab · 2026-09-23	AIME-Plus-Plus (AIME++) Other	“Finally, we plot models’ performance with filler tokens on two sets of math competition problems: AIME-Plus-Plus (AIME++) : AIME-level or higher problems with altered numbers to prevent contamination.” source
Claude Opus 5 Anthropic	Redwood Research Independent lab · 2026-09-23	AIME/HMMT Other	“AIME/HMMT : a dataset of AIME and HMMT questions from 2024-26, gathered from MathArena .” source
Claude Opus 4.5 Anthropic	Redwood Research Independent lab · 2026-09-23	n-hop natural facts Other	“On tasks designed to require lots of serial cognition, Astra performs significantly better with filler tokens than without (e.g., improving from ~10-20% 2 to ~50% on 4-hop natural facts reasoning).” source
Claude Opus 4.5 Anthropic	Redwood Research Independent lab · 2026-09-23	gen-arithmetic Other	“Takeaway : Astra improves much more on these problems than other models.” source
Claude Opus 4.5 Anthropic	Redwood Research Independent lab · 2026-09-23	aime-plus-plus (aime++) Other	“Takeaway : Astra again improves on AIME-level questions with more filler tokens, while other models do not.” source
Claude Opus 4.5 Anthropic	Redwood Research Independent lab · 2026-09-23	humanity’s last exam (hle) Other	“Takeaway: Filler tokens moderately improve most HLE and LiveBench categories.” source
Claude Opus 4.5 Anthropic	Redwood Research Independent lab · 2026-09-23	kendall's tau test Other	“Takeaway: Besides Astra, only GPT-5.6-Sol shows improvement on three of the benchmarks we test.” source
Claude Opus 4.5 Anthropic	Epoch AI Independent lab · 2026-09-23	Furniture Assembly Benchmark (FAB) Other	“The best score in November 2025 was 28%, set by Anthropic’s Claude Opus 4.5.” source
Claude Fable 5.1 Anthropic	Epoch AI Independent lab · 2026-09-23	Furniture Assembly Benchmark (FAB) Other	“The highest-scoring model is GPT-6 Astra at 80%, with the next best models, Claude Fable 5.1 and Claude Opus 5, scoring 70% and 61%, respectively.” source
GPT-5.6 Sol OpenAI	Academic Academic · 2026-09-22	BFCL V4 Prompt injection and agents	“In the primary analysis of 159 multi-turn BFCL V4 tasks with complete exact-replay pairs, the complete policy raises task success for GPT-5.6 Sol from 45.3% to 58.5% (95% task-bootstrap CI [8.2, 18.8]), with no observed success-to-failure regressions.” source
GPT-5.6 Luna OpenAI	Academic Academic · 2026-09-22	Terminal-Bench 2.1 Other	“Across the complete 87-task Terminal-Bench 2.1 suite, with two attempts per task, policies increase repeated success (pass^2) in all three GPT-5.6 tiers: 50.6% to 54.0% for Luna, 55.2% to 60.9% for Terra, and 64.4% to 73.6% for Sol.” source
GPT-5 OpenAI	Academic Academic · 2026-09-22	leave-one-agent-out calibration audit Other	“Two specific combinations, gpt-5-mini/SUCCESS and claude-opus-4.6/FAILURE, showed persistent calibration-transfer errors (median corrected gaps 0.1377 and 0.1107) without a sign reversal under any frozen control.” source

Previous · Page 8 of 48 · Next