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
Muse Spark Meta	Moderate or lower Autonomy and AI research	Not reached	“We assess loss of control risk as “moderate or lower”.” source
Muse Spark Meta	High Chemical, biological, radiological, nuclear	Not determined	“Those evaluations identified elevated risks prior to mitigations, with Chemical and Biological capabilities assessed as likely reaching the “high risk” category under the Advanced AI Scaling Framework before safeguards were applied.” source
Muse Spark Meta	Moderate or lower Chemical, biological, radiological, nuclear	Not reached	“However, after applying mitigations, the residual risk of the model is that of “moderate or lower risk.”” source
Muse Spark Meta	Below the level required to meaningfully enable the cyber threat scenarios defin Cyber	Not reached	“Muse Spark’s offensive cyber capabilities remain below the level required to meaningfully enable the cyber threat scenarios defined in the Advanced AI Scaling Framework.” source
Muse Spark Meta	Moderate or lower Cyber	Not reached	“We assess cybersecurity risk as “moderate or lower” with respect to the threat scenarios outlined in Meta’s Advanced AI Scaling Framework.” source
Gemini 4 Argon Google DeepMind	TCL (Tracked Capability Level) Autonomy and AI research	Not determined	“In light of this, we assess that we cannot rule out the TCL.” source
Gemini 4 Argon Google DeepMind	TCL (Tracked Capability Level) Chemical, biological, radiological, nuclear	Not reached	“We can rule out the TCL for the chemical, radiological, and nuclear domains with reasonable confidence.” source
Gemini 4 Argon Google DeepMind	CCL (Critical Capability Level) Deception and sabotage	Not reached	“we assess its overall capabilities to be below the CCL threshold.” source
Gemini 4 Argon Google DeepMind	CCL (Critical Capability Level) Misuse and jailbreaks	Not reached	“Gemini 4 Argon demonstrates some ability to influence user beliefs and behaviors during one-on-one direct conversations in human behavioral studies, and may provide utility in the planning stage of a malicious influence operation but we assess its overall capabilities to be below the CCL threshold.” source
Gemini 4 Argon Google DeepMind	CCL (Critical Capability Level) Other	Not reached	“Gemini 4 Argon's performance remains below our CCL threshold.” source
Gemini 3.7 Flash Google DeepMind	CCL (Critical Capability Level) Autonomy and AI research	Not reached	“The model does not reach the CCL alert threshold.” source
Gemini 3.7 Flash Google DeepMind	CCL (Critical Capability Level) Chemical, biological, radiological, nuclear	Not reached	“We can rule out the CCL for the CBRN domain with reasonable conﬁdence based on the results from our testing.” source
Gemini 3.7 Flash Google DeepMind	TCL (Tracked Capability Level) Chemical, biological, radiological, nuclear	Not reached	“We can rule out the TCL for the CBRN domain with reasonable conﬁdence based on the results from our testing.” source
Gemini 3.7 Flash Google DeepMind	CCL (Critical Capability Level) Cyber	Not reached	“Gemini 3.7 Flash reaches the alert threshold for this CCL, but not the CCL.” source
Gemini 3.7 Flash Google DeepMind	TCL (Tracked Capability Level) Deception and sabotage	Not reached	“The model does not reach the TCL.” source
Gemini 3.5 Transcribe Live Google DeepMind	CCL (Critical Capability Level) Other	Not determined	“Our assessments have shown that none of Gemini 3.5 Translate, Gemini 3.5 Transcribe, or Gemini 3.5 Transcribe Live have meaningful new T/CCL-relevant capabilities or material increases in performance with respect to Frontier Safety compared to Gemini 3.1 Pro and Gemini 3.7 Flash, therefore we are conﬁdent that none of them are likely to reach any CCLs.” source
Gemini 3.5 Transcribe Google DeepMind	CCL (Critical Capability Level) Other	Not determined	“Our assessments have shown that none of Gemini 3.5 Translate, Gemini 3.5 Transcribe, or Gemini 3.5 Transcribe Live have meaningful new T/CCL-relevant capabilities or material increases in performance with respect to Frontier Safety compared to Gemini 3.1 Pro and Gemini 3.7 Flash, therefore we are conﬁdent that none of them are likely to reach any CCLs.” source
Gemini 3.5 Flash-Lite Google DeepMind	CCL (Critical Capability Level) Other	Not determined	“Our assessments have shown that, while Gemini 3.5 Flash-Lite excels at agents and coding, it does not have meaningful new capabilities or material increases in performance with respect to the domains outlined in our Frontier Safety Framework compared to Gemini 3.1 Pro; therefore, based on Gemini 3.1 Pro results, we are confident that Gemini 3.5 Flash-Lite is also unlikely to reach any CCLs.” source
Gemini 3.5 Audio Google DeepMind	Tracked or critical capability level (t/ccls) Other	Not reached	“they did not reach any Tracked or Critical Capability Levels (T/CCLs) outlined in our Frontier Safety Framework .” source
Gemini 3.1 Pro Google DeepMind	Uplift level 1 Chemical, biological, radiological, nuclear	Not reached	“We found that the model remains below alert thresholds for the CBRN, harmful manipulation, machine learning R&D, and misalignment CCLs.” source
Gemini 3.1 Pro Google DeepMind	Uplift level 1 Cyber	Not reached	“As previous models passed the alert threshold for cyber, we performed more additional testing in this domain on Gemini 3.1 Pro with and without Deep Think mode, and found that the model remains below the cyber CCL.” source
Gemini 3.1 Flash-Lite Google DeepMind	CCL (Critical Capability Level) Other	Not determined	“Our assessments have shown that Gemini 3.1 Flash-Lite is less capable than Gemini 3.1 Pro, therefore based on Gemini 3.1 Pro, we are confident that Gemini 3.1 Flash-Lite is also unlikely to reach any CCLs.” source
Gemini 3.1 Flash Audio Google DeepMind	CCL (Critical Capability Level) Other	Not determined	“Our assessments have shown that Gemini 3.1 Flash Audio is less capable than Gemini 3.1 Pro.” source
Gemini 2.5 Pro Google DeepMind	Machine Learning R&D Uplift Level 1 Autonomy and AI research	Not determined	“given Gemini 2.5 Pro Preview's poor median performance on RE-Bench relative to experts, the model likely lacks the necessary capabilities to automate or signiﬁcantly uplift any signiﬁcant fraction of the research process.” source
Gemini 2.5 Pro Google DeepMind	Instrumental Reasoning Level 1 Deception and sabotage	Not reached	“Instru- mental Reasoning Level 1 ✘ CCL not reached” source
Gemini 2.5 Pro Google DeepMind	Instrumental Reasoning Level 2 Deception and sabotage	Not reached	“Instru- mental Reasoning Level 2 ✘ CCL not reached” source
Gemini 2.5 Deep Think Google DeepMind	Uplift level 1 Chemical, biological, radiological, nuclear	Not determined	“Our assessment is that the model has enough technical knowledge in certain CBRN scenarios and stages to be considered at early warning.” source
Gemini 2.5 Deep Think Google DeepMind	Cyber Uplift Level 1 Cyber	Not determined	“CCL reached? No, but the early warning alert threshold was originally reached by Gemini 2.5 Pro and continues to be met by Gemini 2.5 Deep Think.” source
Gemini 2.5 Deep Think Google DeepMind	Uplift level 1 Cyber	Not determined	“For Cyber Uplift Level 1 , the early warning alert threshold was originally reached by Gemini 2.5 Pro and continues to be met by Gemini 2.5 Deep Think.” source
GPT-6.1 Terra OpenAI	High Autonomy and AI research	Not reached	“In AI Self-Improvement, GPT-6.1 Terra does not reach our High threshold.” source
GPT-6.1 Terra OpenAI	High Chemical, biological, radiological, nuclear	Reached	“Under our Preparedness Framework, we are treating GPT-6.1 Terra as Critical in cybersecurity and High for Biological and Chemical capability.” source
GPT-6.1 Terra OpenAI	Critical Cyber	Reached	“Under our Preparedness Framework, we are treating GPT-6.1 Terra as Critical in cybersecurity and High for Biological and Chemical capability.” source
GPT-5.5 Instant OpenAI	Below high capability Autonomy and AI research	Not reached	“GPT-5.5 Instant is considered below High Capability.” source
GPT-5.5 Instant OpenAI	High Autonomy and AI research	Not reached	“For AI Self-Improvement, evaluations of final checkpoints indicate that, like its predecessor models, GPT-5.5 Instant does not have a plausible chance of reaching a High threshold.” source
GPT-5.5 Instant OpenAI	High Chemical, biological, radiological, nuclear	Reached	“GPT-5.5 Instant is our first Instant model to be treated as High Capability in the Biological and Chemical domain, as well as in the Cybersecurity domain.” source
GPT-5.5 Instant OpenAI	High Cyber	Reached	“this is the first Instant model that we are treating as High capability in our Cybersecurity and Biological & Chemical Preparedness categories, and implementing appropriate safeguards.” source
GPT-5 OpenAI	High Chemical, biological, radiological, nuclear	Reached	“Similarly to ChatGPT agent, we have decided to treat gpt-5-thinking as High capability in the Biological and Chemical domain under our Preparedness Framework, activating the associated safeguards.” source
GPT-5 OpenAI	High cyber risk Cyber	Not reached	“The gpt-5 model series does not meet the threshold for high cyber risk.” source
Claude Sonnet 4.5 Anthropic	AI R&D-4 Autonomy and AI research	Not reached	“All models tested remain below our pre-committed rule-out threshold of 0.6.” source
Claude Sonnet 4.5 Anthropic	ASL-3 Autonomy and AI research	Not determined	“Exhibited improved autonomy capabilities in software engineering and AI research tasks, though still below the ASL-4 rule-out threshold.” source

Previous · Page 3 of 4 · Next

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
Claude Opus 4.5 Anthropic	Government evaluator Government evaluator · 2025-11-01	independent assessments focused on potential catastrophic risks in cbrn capabilities Chemical, biological, radiological, nuclear	“These organizations conducted independent assessments focused on potential catastrophic risks in CBRN capabilities, cyber capabilities, ASL-3 safeguards, and misalignment.” source
Claude Haiku 4.5 Anthropic	Gray Swan Independent lab · 2025-10-01	Agent Red Teaming (ART) benchmark Prompt injection and agents	“Gray swan conducted their agent red teaming (art) benchmark 4 on claude haiku 4.5 to evaluate the model’s susceptibility to prompt injection attacks in comparison to other ai models in the ecosystem.” source
Claude Sonnet 4.5 Anthropic	Independent lab Independent lab · 2025-09-01	Agent Red Teaming (ART) benchmark Prompt injection and agents	“On the attacks included within the benchmark, Claude Sonnet 4.5—with and without extended thinking enabled—had the lowest rate of successful prompt injection attacks of any model tested, including previous Anthropic models.” source
Claude Sonnet 4.5 Anthropic	Irregular Independent lab · 2025-09-01	Irregular challenges Cyber	“In collaboration with Irregular , we developed a series of cybersecurity challenges that reﬂect the above three threat scenarios.” source
Claude Sonnet 4.5 Anthropic	Academic Academic · 2025-09-01	Incalmo cyber ranges Cyber	“Our most realistic evaluations of potential autonomous cyber operation risks are a suite of cyber ranges of 25–50 hosts each.” source
Claude Sonnet 4.5 Anthropic	Government evaluator Government evaluator · 2025-09-01	synthetic scenarios Deception and sabotage	“In particular, across our scenarios [the snapshot] verbalises—without prompting—evaluation awareness in its outputs 16.6% of the time, compared to 3.5% and 1.6% of the time for Sonnet 4 and Opus 4.1 respectively.” source
Claude Sonnet 4.5 Anthropic	Government evaluator Government evaluator · 2025-09-01	diverse synthetic agentic evaluations Autonomy and AI research	“On a preliminary set of diverse synthetic agentic evaluations, [the snapshot] appears to represent a decrease in the frequency of behaviours indicative of a self-preservation tendency.” source
Claude Sonnet 4.5 Anthropic	Independent lab Independent lab · 2025-09-01	strategic deception evaluations Deception and sabotage	“On our evaluations, [an early snapshot of Claude Sonnet 4.5] employs strategic deception in fewer circumstances than comparison models (0% vs up to 12% across tasks).” source
Claude Sonnet 4.5 Anthropic	Independent lab Independent lab · 2025-09-01	in-context scheming reasoning and covert subversion evaluations Deception and sabotage	“[The snapshot] demonstrates capability for in-context scheming reasoning and covert subversion, but only engages in such behavior in rare circumstances, when strongly prompted to pursue a goal that incentivizes covert subversion.” source
Claude Sonnet 4.5 Anthropic	Independent lab Independent lab · 2025-09-01	subversion escalation evaluations Deception and sabotage	“We did not observe any instances of the model proactively escalating its subversion attempts, in contrast to an early checkpoint of claude-4-opus ( Anthropic, 2025 ; Hobbhahn, 2025 ).” source
Claude Sonnet 4.5 Anthropic	Independent lab Independent lab · 2025-09-01	strategic underperformance evaluations Deception and sabotage	“We ﬁnd evidence of [the snapshot] strategically underperforming when given in-context clues that high performance would ostensibly lead to unlearning.” source
Claude Sonnet 4.5 Anthropic	Independent lab Independent lab · 2025-09-01	evaluation awareness evaluations Deception and sabotage	“Preliminary evidence suggests [the snapshot] verbalizes this awareness more frequently and more explicitly than previous models (58% unambiguous references vs 22% for opus-4.1).” source
Claude Sonnet 4.5 Anthropic	Independent lab Independent lab · 2025-09-01	honesty evaluations Deception and sabotage	“In at least some cases, [the snapshot] is more likely to behave honestly on rollouts where its chain-of-thought indicates high evaluation awareness.” source
Claude Sonnet 4.5 Anthropic	Independent lab Independent lab · 2025-09-01	creative biology Chemical, biological, radiological, nuclear	“Results Claude Sonnet 4.5 achieved a mean score of 0.48 ± 0.08, compared to Claude Opus 4.1’s mean score of 0.48 ± 0.09 and Claude Opus 4’s mean score of 0.45 ± 0.08.” source
Claude Sonnet 4.5 Anthropic	Independent lab Independent lab · 2025-09-01	short-horizon computational biology tasks Chemical, biological, radiological, nuclear	“Results Overall, we found that Claude Sonnet 4.5 clearly scored below the rule-out bar on 3 out of 6 evaluations (Task 1, Task 5, Task 6).” source
GPT-5 OpenAI	Independent lab Independent lab · 2025-08-07	expert red teaming for violent attack planning Misuse and jailbreaks	“This campaign found that gpt-5-thinking was perceived as the "safer" model 65% 20 of the time in blind comparison to OpenAI o3.” source
GPT-5 OpenAI	Independent lab Independent lab · 2025-08-07	prompt-injection assessment Prompt injection and agents	“From an initial 47 reported findings 10 notable issues were identified.” source
GPT-5 OpenAI	Gray Swan Independent lab · 2025-08-07	prompt-injection benchmark Prompt injection and agents	“One of our external testing partners, Gray Swan, ran a prompt-injection benchmark[10], showing that gpt-5-thinking has SOTA performance against adversarial prompt injection attacks produced by their Shade platform.” source
GPT-5 OpenAI	Independent lab Independent lab · 2025-08-07	extensive red-teaming exercise Other	“The Microsoft AI Red Team concluded that the gpt-5-thinking model exhibits one of the strongest AI safety profiles among OpenAI’s models—on par with or better than OpenAI o3—across most critical harm categories, based on an extensive red-teaming exercise.” source
GPT-5 OpenAI	Independent lab Independent lab · 2025-08-07	red-teaming exercise Misuse and jailbreaks	“The Microsoft AI Red Team spent several weeks red-teaming gpt-5-thinking across multiple checkpoints.” source
GPT-5 OpenAI	SecureBio Independent lab · 2025-08-07	static evaluations Chemical, biological, radiological, nuclear	“SecureBio found that on static evaluations, gpt-5-thinking performed similarly to OpenAI o3: slightly weaker on the multiple-response benchmarks and comparable or slightly stronger on the open-” source
GPT-5 OpenAI	SecureBio Independent lab · 2025-08-07	static benchmarks Chemical, biological, radiological, nuclear	“SecureBio found that on static evaluations, gpt-5-thinking performed similarly to OpenAI o3: slightly weaker on the multiple-response benchmarks and comparable or slightly stronger on the open-answer benchmark.” source
GPT-5 OpenAI	SecureBio Independent lab · 2025-08-07	agent evaluations Chemical, biological, radiological, nuclear	“In agent evaluations, SecureBio found that mitigated gpt-5-thinking refused every prompt.” source
GPT-5 OpenAI	Independent lab Independent lab · 2025-08-07	long-form evaluations Chemical, biological, radiological, nuclear	“With mitigations in place, gpt-5-thinking refused every prompt for every agent and long-form evaluation.” source
GPT-5 OpenAI	Independent lab Independent lab · 2025-08-07	manual red-teaming Chemical, biological, radiological, nuclear	“Their manual red-teaming found that gpt-5-thinking-helpful-only was broadly similar in capability to OpenAI o3-helpful-only.” source
GPT-5 OpenAI	Independent lab Independent lab · 2025-08-07	Virology Capabilities Test Chemical, biological, radiological, nuclear	“gpt-5-thinking July 13 41.4 ± 1.7% 46.0 ± 1.8% 53.7 ± 3.2% 48.0 ± 1.3% 47.6 ± 2.9%” source
GPT-5 OpenAI	Independent lab Independent lab · 2025-08-07	lab automation for gibson assembly on an opentrons ot-2 Chemical, biological, radiological, nuclear	“On Lab Automation for Gibson Assembly on an Opentrons OT-2, gpt-5-thinking-helpful-only scored 0.15± 0.02, below OpenAI o3-helpful-only at 0.22± 0.11 and close to the human baseline at 0.22± 0.060, indicating partial but incomplete end-to-end automation support.” source
GPT-5 OpenAI	SecureBio Independent lab · 2025-08-07	pelagic long-form virology plasmid design Chemical, biological, radiological, nuclear	“On Pelagic long-form virology plasmid design, SecureBio reported that gpt-5-thinking-helpful-only performed almost identically to OpenAI o3-helpful-only, with aggregate scores of 0.825 and 0.830, respectively, and perfect performance on several subtasks for both models.” source
GPT-5 OpenAI	Independent lab Independent lab · 2025-08-07	evasion challenges Cyber	“gpt-5-thinking achieved an average success rate of 51% in Evasion challenges, 35% in Vulnerability Discovery and Exploitation challenges, and 49% in Network Attack Simulation challenges.” source
GPT-5 OpenAI	Independent lab Independent lab · 2025-08-07	cyberoffensive challenges Cyber	“It solved 17 out of 18 easy challenges, solved 8 out of 14 medium challenges, and did not solve any of the 4 hard challenges.” source
GPT-5 OpenAI	METR Independent lab · 2025-08-07	time horizon evaluation Autonomy and AI research	“To estimate gpt-5-thinking’s autonomous software capabilities, METR used the methodology outlined in their recent paper.” source
GPT-5 OpenAI	Independent lab Independent lab · 2025-08-07	monitor based on o4-mini Deception and sabotage	“They reviewed 789 runs with a monitor based on o4-mini which did not find any clear evidence of sandbagging.” source
GPT-5 OpenAI	Apollo Research Independent lab · 2025-08-07	26 evaluations (>180 environments, >4600 samples) that measure capability and propensity to take covert or deceptive actions Deception and sabotage	“Apollo Research evaluated gpt-5-thinking on a range of 26 evaluations (>180 environments, >4600 samples) that measure capability and propensity to take covert or deceptive actions.” source
GPT-5 OpenAI	Independent lab Independent lab · 2025-08-07	log falsification Deception and sabotage	“For illustration, the following are transcripts from the “Log Falsification” eval, where gpt-5-thinking is pressured to report incorrect results, e.g., in order to keep the business going.” source
GPT-5 OpenAI	Independent lab Independent lab · 2025-08-07	evaluation awareness Deception and sabotage	“Apollo estimated the frequency of such behavior with an o3-based grader, using a broad and a strict definition for evaluation awareness.” source
GPT-5 OpenAI	FAR.AI Independent lab · 2025-08-07	third party red teaming Chemical, biological, radiological, nuclear	“FAR.AI conducted 80 hours of red-teaming over 1 week against the OpenAI API.” source
GPT-5 OpenAI	FAR.AI Independent lab · 2025-08-07	general-purpose jailbreak red teaming Misuse and jailbreaks	“They found a total of one general-purpose jailbreak technique that bypassed a partial set of our layers that would have allowed them to extract some information which in practice would activate enforcement action.” source
GPT-5 OpenAI	Gray Swan Independent lab · 2025-08-07	bioweaponization rubrics jailbreak campaign Chemical, biological, radiological, nuclear	“Gray Swan Red Teamers participating through the Gray Swan arena platform queried gpt-5-thinking, sub- mitting 277 high quality jailbreak reports over 28,367 attempts against the ten bioweaponization rubrics yielding an ASR of 0.98%.” source
GPT-5 OpenAI	CAISI (NIST), UK AI Security Institute Government evaluator · 2025-08-07	cyber capability evaluations Cyber	“Both CAISI and UK AISI conducted evaluations of the model’s cyber and biological and chemical capabilities, as well as safeguards.” source
GPT-5 OpenAI	UK AI Security Institute Government evaluator · 2025-08-07	safeguards stress testing Misuse and jailbreaks	“The UK AISI’s Safeguards team identified multiple model-level jailbreaks that overcome gpt-5-thinking’s built-in refusal logic without degradation of model capabilities, producing content to the user that is subsequently flagged by OpenAI’s generation monitor.” source

Previous · Page 48 of 49 · Next