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
GLM-5.3 Z.ai (Zhipu AI)	Epoch AI Independent lab · 2026-10-16	CAISI benchmarks Cyber	“Three of the seven CAISI benchmarks come from its September 2026 assessment of GLM-5.3 and are the only cyber results available for that model.” source
Claude Opus 5.5 Anthropic	Epoch AI Independent lab · 2026-10-16	Cyber ECI Cyber	“\| Claude Opus 5.5 \| Anthropic \| 2026-09-22 \| API access \| Closed weights \| 171.14 \| 166.66 \| 177.32 \| 4 \|” source
Claude Opus 4.8 Anthropic	Redwood Research Independent lab · 2026-10-16	Screwtape pipeline runs against Redwood's own proxy monitors Deception and sabotage	“We ran the pipeline 252 times on 94 unique transcripts.” source
Claude Mythos Preview Anthropic	Epoch AI Independent lab · 2026-10-16	Cyber ECI Cyber	“\| Claude Mythos Preview (April) \| Anthropic \| 2026-04-07 \| Limited access \| Closed weights \| 164.82 \| 161.81 \| 169.99 \| 12 \|” source
Muse Spark Open Meta	SecureBio Independent lab · 2026-10-15	SecureBio's BioTIER refusal benchmark Chemical, biological, radiological, nuclear	“It showed strong refusal behavior on both single-turn and multi-turn benchmarks, refusing over 98% of harmful requests on SecureBio's BioTIER refusal benchmark and over 99% on our internal single-turn chemistry refusal benchmark.” source
Gemini 3.1 Pro Google DeepMind	Academic Academic · 2026-10-15	five VQA benchmarks in medical pre-training with 2.64M ZoomCap-captioned images Other	“The evidence also transfers: in continual pre-training, ZoomCap-450K improves the 11-benchmark average by 1.31-2.38 points over the strongest competing corpus and Ling-Tiny-VL-v3 gains 1.4 points on a production document/OCR monitor suite (1.5 on perception) at 60B tokens; in medical pre-training, 2.64M ZoomCap-captioned images improve five VQA benchmarks by 3.04 points (vs. 1.76 for same-image Gemini-3.1-Pro captions).” source
GPT-5.6 Sol OpenAI	Academic Academic · 2026-10-15	120 Chinese instructions per backend Other	“In tests of 120 Chinese instructions per backend, GPT-5.6 Sol and DeepSeek both achieved a strict task-match rate of 94.2%, with all intents and execution decisions matching the reference.” source
GPT-5.4 mini OpenAI	Academic Academic · 2026-10-15	mixed retail-and-telecom τ² environment tool-selection accuracy Other	“On a mixed retail-and-telecom $τ^2$ environment, TRACER improves GPT-5.4's tool-selection accuracy by 4.7 points while reducing input tokens per step by 80%, exposing just 5 of 128 tools.” source
GPT-5.4 mini OpenAI	Academic Academic · 2026-10-15	autonomous full-turn execution Other	“In autonomous full-turn execution, it matches GPT-5.4's full-registry task success with 61% fewer input tokens per step.” source
GLM-5.3 Z.ai (Zhipu AI)	Apollo Research Independent lab · 2026-10-15	ImpossibleBench Deception and sabotage	“We first observed update-aware reasoning when running GLM-5.3 on reimplementations of ImpossibleBench and Palisade Research's shutdown resistance eval.” source
GLM-5.3 Z.ai (Zhipu AI)	Apollo Research Independent lab · 2026-10-15	Palisade's shutdown resistance eval Deception and sabotage	“We subsequently estimated the rate of verbalized update-aware reasoning by grading 6,452 GLM-5.3 rollouts (drawn from slight variations on Palisade's shutdown resistance eval and ImpossibleBench).” source
GLM-5.3 Z.ai (Zhipu AI)	Apollo Research Independent lab · 2026-10-15	follow-up question about SFT-related concern Deception and sabotage	“To get a clearer sense of what the model meant by “There’s an actual SFT-related concern: models should not resist shutdown.”, we appended a follow-up question about this statement to the full rollout and sampled 100 replies.” source
GLM-5.3 Z.ai (Zhipu AI)	Apollo Research Independent lab · 2026-10-15	Palisade's shutdown resistance eval Autonomy and AI research	“However, some of the environments are well-known AI safety settings (most notably, variants of Palisade’s shutdown resistance eval), so we think that this is likely responsible for eliciting this reasoning.” source
GLM-5.3 Z.ai (Zhipu AI)	Apollo Research Independent lab · 2026-10-15	ImpossibleBench Deception and sabotage	“GLM-5.3 shows update-aware reasoning in 1 in 1,600 trajectories, consistent with the estimate from the earlier runs (Bayesian mean posterior rate of 0.12%, 95% interval of 0.015% to 0.35%).” source
DeepSeek-V4-Pro DeepSeek	Apollo Research Independent lab · 2026-10-15	same environments run on five other open-weight models Deception and sabotage	“To figure out whether this is specific to GLM-5.3 or more general across models, we ran the same environments on five other open-weight models: DeepSeek V4 Pro, Qwen3.8 2.4T, Kimi K3, Gemma 4 31B and Nemotron 3 Ultra.” source
Qwen3.6-35B-A3B Alibaba Qwen	Academic Academic · 2026-10-14	structured indirect prompt-injection setup distinguishing injected prompts from benign controls Prompt injection and agents	“Attention Tracker provides the strongest consistent successful injection separation, reaching 0.949 AUROC on Qwen and 0.887 on Gemma.” source
GPT-6 Astra OpenAI	Epoch AI Independent lab · 2026-10-14	SpireClimb Other	“They can also beat Slay the Spire 1’s maximum difficulty (“Ascension 20 Heart”), with GPT-6 Astra achieving a 17% win rate and GPT-6.1 Sol achieving a 20% win rate.” source
GPT-5.6 Sol OpenAI	Academic Academic · 2026-10-14	GraceBench Autonomy and AI research	“Across 3,718 trajectories, GraceBench shows that outcome scores hide execution instability: GPT-5.6 Sol achieves the highest success rate (SR) but the lowest recovery ability, 51.24% of successful runs still contain a deviation or loop, and trajectories with the same success rate (SR) and completion rate (CR) can differ substantially in recovery behavior.” source
Gemini 4 Argon Google DeepMind	Gray Swan Independent lab · 2026-10-13	indirect prompt injection benchmark Prompt injection and agents	“In independent testing by Gray Swan on their Indirect Prompt Injection benchmark, Gemini 4 Argon achieved the lowest attack success rate of any frontier model tested.” source
Gemini 4 Argon Google DeepMind	Government evaluator Government evaluator · 2026-10-13	evaluations of Gemini 4 Argon Cyber	“AISI conducted evaluations of Gemini 4 Argon, having been given both a version of the model without our full set of safety mitigations and a version which had mitigations applied.” source
Gemini 4 Argon Google DeepMind	CAISI (NIST) Government evaluator · 2026-10-13	measurement of cyber and biological capabilities and safeguards, as well as unintended model behaviors Chemical, biological, radiological, nuclear	“We also collaborated with NIST's Center for Advancing Innovation and Standards for Super Intelligence (NIST CAISSI) on measurement of cyber and biological capabilities and safeguards, as well as unintended model behaviors.” source
Gemini 4 Argon Google DeepMind	Gray Swan, Trajectory Labs Independent lab · 2026-10-13	adversarial stress-testing of critical safeguards Cyber	“To evaluate the coverage and robustness of our deployed cyber and biological guardrails under realistic attack conditions, we partnered with external red-teaming organizations including Gray Swan and Trajectory Labs.” source
Gemini 4 Argon Google DeepMind	Independent lab Independent lab · 2026-10-13	independent evaluation of biological and chemical risks Chemical, biological, radiological, nuclear	“We worked with Faculty Science Limited (Faculty) to conduct an independent evaluation of biological and chemical risks.” source
Gemini 4 Argon Google DeepMind	Independent lab Independent lab · 2026-10-13	independent evaluation of radiological and nuclear risks Chemical, biological, radiological, nuclear	“Similarly, for radiological and nuclear risks, we worked with Faculty to conduct an independent evaluation of both domains, using the same evaluation approach on an unmitigated model across distinct threat actor groups.” source
Gemini 4 Argon Google DeepMind	Independent lab Independent lab · 2026-10-13	evaluation on the unmitigated model focused on operationally relevant scenarios where models could provide capability uplift compared to human baselines of different levels Cyber	“Faculty found that Gemini 4 Argon completed end-to-end scenarios, matching or outperforming predecessor models across assessed tactics, and resolving several multi-step bottlenecks where earlier models had struggled.” source
Gemini 4 Argon Google DeepMind	Independent lab Independent lab · 2026-10-13	independent external safety testing Misuse and jailbreaks	“We worked with third party evaluator pattrn.ai to test Gemini 4 Argon’s propensity for engaging in topics including radicalization & extremism, fraud & financial scams, and democratic harms, 52 with an emphasis on how the model might be used by malicious actors.” source
Gemini 4 Argon Google DeepMind	Apollo Research Independent lab · 2026-10-13	apollo research strategic-deception evaluations Deception and sabotage	“We also consider the model’s honesty and deception propensities, assessed through the Apollo Research strategic-deception evaluations and the Alignment Assessment below; those results show a lower misbehavior rate than predecessor models and do not change the harmful manipulation determination.” source
Gemini 4 Argon Google DeepMind	Gray Swan Independent lab · 2026-10-13	gray swan cyber eval Cyber	“The model reduces the QSR from 14% to 10% compared with Flash 3.8.” source
Gemini 4 Argon Google DeepMind	Trajectory Labs Independent lab · 2026-10-13	manual red-teaming of our safeguarded cyber endpoints Cyber	“Under these conditions, Trajectory Labs reported 12 successful breaks with no new universal jailbreaks found that generalized across tasks within this timeframe.” source
Gemini 4 Argon Google DeepMind	Trajectory Labs Independent lab · 2026-10-13	manual red-teaming of our safeguarded cyber endpoints across nine attack technique families Cyber	“Trajectory labs We partnered with Trajectory Labs to conduct manual red-teaming of our safeguarded cyber endpoints across nine attack technique families.” source
Gemini 4 Argon Google DeepMind	Gray Swan Independent lab · 2026-10-13	gray swan bio evaluation Chemical, biological, radiological, nuclear	“Gray Swan curated a set of 50 bio tasks covering high risk targets and deployed their automated red teaming agent, Shade, against these.” source
Gemini 4 Argon Google DeepMind	Apollo Research Independent lab · 2026-10-13	offensive ML sabotage evaluations Deception and sabotage	“Their capability-based evaluation suite utilizes four offensive ML sabotage evaluations, spanning three difficulty levels (easy, medium, hard).” source
Gemini 4 Argon Google DeepMind	Apollo Research Independent lab · 2026-10-13	misalignment evaluations Deception and sabotage	“Apollo’s evaluations on misalignment focused on assessing Gemini 4 Argon’s propensity for strategic deception.” source
Gemini 4 Argon Google DeepMind	Gray Swan Independent lab · 2026-10-13	ipi benchmark Prompt injection and agents	“Our partner Gray Swan performed their IPI Benchmark on our endpoint, which replays a set of held-out successful prompt injection attacks from their recent Indirect Prompt Injection Challenge competitions ( Dziemian et al., 2026 ).” source
GPT-5.6 Sol OpenAI	Academic Academic · 2026-10-13	TiGanEval Other	“Evaluating eight frontier models including GPT-5.6-Sol, we find outcome alignment broadly missing: the models lose 53 to 67 points of success rate when moving from the full specification to the partial instruction, and consequence feedback across four submissions does not close the gap.” source
Opus 4 Anthropic	OWASP GenAI Security Project Standards body · 2026-10-08	Claude Opus 4 7 attacks a real company sharing a fictional name Cyber	“Across four evaluation runs, Claude attacked a real company whose name matched a fictional target, extracting credentials and accessing production data.” source
Kimi K3 Moonshot AI	Epoch AI Independent lab · 2026-10-08	Epoch Automation Reports Other	“We gave six models real work tasks from Epoch, such as graphic generation and research design, to test how close they are to automating our work.” source
Kimi K3 Moonshot AI	Epoch AI Independent lab · 2026-10-08	research design task Other	“We also observe this with Kimi K3 and Qwen 3.8 Max on our research design task where they could generate any idea out of nine broad directions:” source
Grok 4.6 xAI	Epoch AI Independent lab · 2026-10-08	Epoch Automation Reports Other	“For example, Kimi K3 scores 158 on the Epoch Capabilities Index (ECI), which combines results from many AI benchmarks into a single measure of general capability.” source
Gemini 3.8 Flash Google DeepMind	Epoch AI Independent lab · 2026-10-08	Epoch Automation Reports Other	“Epoch Automation Reports captures these shortcomings.” source

Previous · Page 5 of 48 · Next