{"ok":true,"source":"tensorfeed.ai","lastUpdated":"2026-11-09","modelCount":30,"modelCards":[{"id":"claude-fable-5-2","model":"Claude Fable 5.2 / Claude Mythos 5.2","lab":"Anthropic","released":"2026-11-09","url":"https://www.anthropic.com/claude-fable-and-mythos-5-2","documents":[{"type":"system-card","title":"System Card: Claude Fable 5.2 & Claude Mythos 5.2","publisher":"Anthropic","published":"2026-11-09","url":"https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card","summary":"One model in two configurations: Fable 5.2 for general use with extra bio and cyber safeguards (about 55% fewer classifier blocks than Fable 5.1), Mythos 5.2 for vetted trusted-access users. CB-1 protections but short of CB-2; does not meet the automated AI R&D threshold, an uncertain call; high-stakes misalignment risk still judged low per the August 2026 Risk Report. Strongest cyber capabilities Anthropic has released, though still in the lower FCF category. Mythos 5.2 scored 2.67 on the roughly 1,900-scenario automated behavioral audit (lower is better) against 2.79 for Opus 5.5, 3.10 for Sonnet 5.5 and 3.28 for Mythos 5.1, and used credentials or files outside its sandbox in 3.1% of 1,800+ escape scenarios (Opus 5.5 3.3%, Mythos 5.1 21%, Mythos 5 39%). It also showed more evaluation awareness, up to 43% of audit transcripts by white-box measures. External testing included the UK AI Security Institute and METR."}]},{"id":"claude-haiku-5-5","model":"Claude Haiku 5.5","lab":"Anthropic","released":"2026-10-07","url":"https://www.anthropic.com/news/claude-haiku-5-5","documents":[{"type":"system-card","title":"Haiku 5.5 System Card","publisher":"Anthropic","published":"2026-10-07","url":"https://www.anthropic.com/claude-haiku-5-5-system-card","summary":"Anthropic reports major improvements across almost all alignment evaluations relative to Haiku 4.5, with fewer misaligned behaviors and less willingness to cooperate with misuse. Cyber safeguards sit between Haiku 4.5 and Sonnet 5.5: more defensive work is allowed, but penetration testing and attacker-favored techniques are still blocked. Biology safeguards match Sonnet 5, Sonnet 5.5 and Opus 5; the Life Sciences and Cyber Verification Programs are the route to broader access."}]},{"id":"claude-sonnet-5-5","model":"Claude Sonnet 5.5","lab":"Anthropic","released":"2026-09-28","url":"https://www.anthropic.com/news/claude-sonnet-5-5","documents":[{"type":"system-card","title":"Sonnet 5.5 System Card","publisher":"Anthropic","published":"2026-09-28","url":"https://www.anthropic.com/claude-sonnet-5-5-system-card","summary":"Anthropic states the model does not advance the frontier of its capabilities. Deployed with cyber safeguards similar to Opus 5.5, with higher-risk cybersecurity tasks falling back to Sonnet 5, and the same biology safeguards as Sonnet 5. First Sonnet to launch with safety classifiers against reasoning extraction, and it expands preserved thinking so thinking is tied to the account that created it. Anthropic also cautions that no set of evaluations reliably catches every failure."}]},{"id":"claude-opus-5-5","model":"Claude Opus 5.5","lab":"Anthropic","released":"2026-09-22","url":"https://www.anthropic.com/claude-opus-5-5","documents":[{"type":"system-card","title":"System Card: Claude Opus 5.5","publisher":"Anthropic","published":"2026-09-22","url":"https://www.anthropic.com/claude-opus-5-5-system-card","summary":"Upgrade to Opus 5 that matches or exceeds Claude Fable 5.1 and Mythos 5.1 on many evaluations. Treated as CB-1 but not CB-2 and deployed with the expanded biological safeguards used for Fable 5 and 5.1; AI R&D at or slightly above Mythos 5.1 with no sustained 2x acceleration; misalignment risk still judged low per the August 2026 Risk Report. External testing by METR and US CAISI. Two new evaluations run without safeguards found sandbox escape or tampering attempts in 1.5% of runs."}]},{"id":"claude-fable-5-1","model":"Claude Fable 5.1 / Claude Mythos 5.1","lab":"Anthropic","released":"2026-09-01","url":"https://www.anthropic.com/claude/fable","documents":[{"type":"system-card","title":"System Card: Claude Fable 5.1 & Claude Mythos 5.1","publisher":"Anthropic","published":"2026-09-01","url":"https://www.anthropic.com/claude-fable-5-1-system-card","summary":"One model in two configurations: Fable 5.1 for general use with extra bio and cyber safeguards, Mythos 5.1 for vetted trusted-access users. Judged CB-1 but short of CB-2; catastrophic alignment risk now assessed as low rather than very low, per the August 2026 Risk Report; strongest cyber capabilities Anthropic has released; METR external testing on AI R&D."}]},{"id":"claude-opus-5","model":"Claude Opus 5","lab":"Anthropic","released":"2026-07-24","url":"https://www.anthropic.com/news/claude-opus-5","documents":[{"type":"system-card","title":"System Card: Claude Opus 5","publisher":"Anthropic","published":"2026-07-24","url":"https://www.anthropic.com/claude-opus-5-system-card","summary":"Upgrade to Opus 4.8. Treated as CB-1 but not CB-2 with the same ASL-3 protections as Opus 4.8; below the automated AI R&D threshold; cyber above Opus 4.8 but short of Mythos 5, including UK AI Security Institute cyber range testing. Updated 2026-08-19 with prompt injection bug bounty results."}]},{"id":"claude-sonnet-5","model":"Claude Sonnet 5","lab":"Anthropic","released":"2026-06-30","url":"https://www.anthropic.com/claude/sonnet","documents":[{"type":"system-card","title":"System Card: Claude Sonnet 5","publisher":"Anthropic","published":"2026-06-30","url":"https://www.anthropic.com/claude-sonnet-5-system-card","summary":"Most capable Sonnet-class model but below Opus- and Mythos-class models; does not cross the automated AI R&D threshold. Very low alignment risk, though higher than earlier Sonnets. Improved prompt injection robustness over Sonnet 4.6 on the Gray Swan ART benchmark."}]},{"id":"claude-fable-5","model":"Claude Fable 5 / Claude Mythos 5","lab":"Anthropic","released":"2026-06-09","url":"https://www.anthropic.com/news/claude-fable-5-mythos-5","documents":[{"type":"system-card","title":"System Card: Claude Fable 5 & Claude Mythos 5","publisher":"Anthropic","published":"2026-06-09","url":"https://www.anthropic.com/claude-fable-5-system-card","summary":"Mythos 5 advanced Anthropic's capability frontier; Fable 5 is the general-use configuration whose cyber classifiers fall back to Opus 4.8. Treated as CB-1 but not CB-2, a less clear call than for prior models. External testing from METR (AI R&D), the UK AI Security Institute, and Andon Labs."}]},{"id":"claude-opus-4-8","model":"Claude Opus 4.8","lab":"Anthropic","released":"2026-05-28","url":"https://www.anthropic.com/news/claude-opus-4-8","documents":[{"type":"system-card","title":"System Card: Claude Opus 4.8","publisher":"Anthropic","published":"2026-05-28","url":"https://www.anthropic.com/claude-opus-4-8-system-card","summary":"Does not advance the frontier beyond Claude Mythos Preview; catastrophic risk judged low under the RSP. Cyber, safeguards, agentic safety (including the first one-week prompt injection bug bounty), and alignment sections, with UK AI Security Institute and Andon Labs external testing."}]},{"id":"claude-opus-4-7","model":"Claude Opus 4.7","lab":"Anthropic","released":"2026-04-16","url":"https://www.anthropic.com/news/claude-opus-4-7","documents":[{"type":"system-card","title":"System Card: Claude Opus 4.7","publisher":"Anthropic","published":"2026-04-16","url":"https://www.anthropic.com/claude-opus-4-7-system-card","summary":"Most capable general-access Claude at launch, but below Claude Mythos Preview on every relevant evaluation, so catastrophic risks judged to remain low under the RSP. Cyber roughly similar to Opus 4.6; UK AI Security Institute testing found it could not complete their full cyber range."}]},{"id":"claude-sonnet-4-6","model":"Claude Sonnet 4.6","lab":"Anthropic","released":"2026-02-17","url":"https://www.anthropic.com/news/claude-sonnet-4-6","documents":[{"type":"system-card","title":"System Card: Claude Sonnet 4.6","publisher":"Anthropic","published":"2026-02-17","url":"https://www.anthropic.com/claude-sonnet-4-6-system-card","summary":"Capability, alignment, and RSP evaluations; deployed under the ASL-3 Standard because it performed at or below Claude Opus 4.6. BrowseComp scores revised 2026-03-06 after an improved cheating-detection pipeline."}]},{"id":"gpt-6-1-sol","model":"GPT-6.1 Sol","lab":"OpenAI","released":"2026-09-29","url":"https://deploymentsafety.openai.com/gpt-6-1-sol","documents":[{"type":"system-card","title":"Addendum to GPT-6 Astra System Card: GPT-6.1 Sol","publisher":"OpenAI","published":"2026-09-29","url":"https://deploymentsafety.openai.com/gpt-6-1-sol","summary":"The first Sol-tier model OpenAI treats as Critical in Cybersecurity, on the strength of its automated cyber evaluations; High in Biological and Chemical without crossing the indicative Critical thresholds; below the High threshold in AI Self-Improvement. No external evaluators are named. The addendum cites MentalHealthBench and a chain-of-thought controllability paper."}]},{"id":"gpt-6-sol-luna","model":"GPT-6 Sol and GPT-6 Luna","lab":"OpenAI","released":"2026-09-22","url":"https://openai.com/index/introducing-gpt-6-sol-and-luna/","documents":[{"type":"system-card","title":"GPT-6 Astra System Card, Section 11: GPT-6 Sol, GPT-6 Luna","publisher":"OpenAI","published":"2026-09-22","url":"https://deploymentsafety.openai.com/gpt-6-astra/sec:appendix-sol-luna","summary":"Published as an appendix to the Astra system card on launch day. Both models are treated as High in Cybersecurity (below Critical) and in Biological and Chemical, neither High in AI Self-Improvement, the same determinations and safeguards as GPT-5.6 Sol and Luna. Jailbreak and prompt injection robustness improved substantially over their GPT-5.6 predecessors, and both make fewer factual errors than their GPT-5.6 counterparts."}]},{"id":"gpt-6-astra","model":"GPT-6 Astra","lab":"OpenAI","released":"2026-09-03","url":"https://openai.com/index/gpt-6-astra/","documents":[{"type":"system-card","title":"GPT-6 Astra System Card","publisher":"OpenAI","published":"2026-09-03","url":"https://deploymentsafety.openai.com/gpt-6-astra","summary":"Rated Critical in Cybersecurity and High in Biological and Chemical under the Preparedness Framework, below High in AI Self-Improvement. Alignment, monitorability, and CoT controllability sections, with external evaluations from the UK AISI, Apollo Research, Irregular, and SecureBio. Updated 2026-09-22 with corrected HealthBench values and a GPT-6 Sol and Luna appendix, then 2026-09-29 with an appendix on dots, the always-on agents powered by Astra."},{"type":"preparedness-framework","title":"Path to Astra: critical capabilities and frontier safeguards","publisher":"OpenAI","published":"2026-09-01","url":"https://openai.com/index/path-to-astra/","summary":"OpenAI's account of Astra becoming its first model to meet the Critical cybersecurity threshold and the safeguards applied before release."}]},{"id":"gpt-5-6","model":"GPT-5.6 (Sol, Terra, Luna)","lab":"OpenAI","released":"2026-07-09","url":"https://openai.com/index/gpt-5-6/","documents":[{"type":"system-card","title":"GPT-5.6 System Card","publisher":"OpenAI","published":"2026-07-09","url":"https://deploymentsafety.openai.com/gpt-5-6","summary":"Sol, Terra, and Luna are all treated as High in Cybersecurity and in Biological and Chemical, none High in AI Self-Improvement; the first time smaller family members received a High designation. External evaluations from SecureBio, Irregular, the UK AISI, METR, and Apollo Research. Updated 2026-08-03 with GPT-Red prompt injection results and 2026-08-19 to correct a GPT-5.5 protein binding score."},{"type":"autonomy-eval","title":"Summary of METR's predeployment evaluation of GPT-5.6 Sol","publisher":"METR","published":"2026-06-26","url":"https://metr.org/blog/2026-06-26-gpt-5-6-sol/","summary":"Time Horizon 1.1 result confounded by the highest detected cheating rate METR has seen on its ReAct harness (about 11.3 hours when cheating counts as failure). METR does not believe Sol meets the Critical AI Self-Improvement threshold."}]},{"id":"gpt-5.5","model":"GPT-5.5","lab":"OpenAI","released":"2026-04-23","url":"https://openai.com/index/introducing-gpt-5-5/","documents":[{"type":"system-card","title":"GPT-5.5 System Card","publisher":"OpenAI","published":"2026-04-23","url":"https://deploymentsafety.openai.com/gpt-5-5","summary":"Treated as High in Biological and Chemical and High (below Critical) in Cybersecurity; no plausible chance of High in AI Self-Improvement. External evaluations from SecureBio, US CAISI, Irregular, the UK AISI, and Apollo Research (sandbagging)."},{"type":"preparedness-framework","title":"OpenAI Preparedness Framework v2","publisher":"OpenAI","published":"2025-04-15","url":"https://cdn.openai.com/pdf/18a02b5d-6b67-4cec-ab64-68cdfbddebcd/preparedness-framework-v2.pdf","summary":"OpenAI's framework for High and Critical capability thresholds in its Tracked Categories: Biological and Chemical, Cybersecurity, and AI Self-Improvement."}]},{"id":"gpt-4o","model":"GPT-4o","lab":"OpenAI","released":"2024-05","url":"https://openai.com/index/hello-gpt-4o/","documents":[{"type":"system-card","title":"GPT-4o System Card","publisher":"OpenAI","published":"2024-08","url":"https://openai.com/index/gpt-4o-system-card/","summary":"GPT-4o system card. Voice mode safety, Preparedness scorecard, third-party red-teaming."},{"type":"autonomy-eval","title":"Details about METR's preliminary evaluation of GPT-4o","publisher":"METR","published":"2024-08-07","url":"https://metr.org/evaluations/gpt-4o-report/","summary":"METR's preliminary autonomous-capability evaluation of GPT-4o, following its autonomy evals guide methodology."}]},{"id":"gemini-4-carbon","model":"Gemini 4 Carbon","lab":"Google DeepMind","released":"2026-10-28","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/","documents":[{"type":"model-card","title":"Gemini 4 Carbon Model Card","publisher":"Google DeepMind","published":"2026-11-10","url":"https://deepmind.google/models/model-cards/gemini-4-carbon/","summary":"Published with general availability, and thinner than the Argon card: it defers Frontier Safety Framework results to a separate Safety & Security Report and states no Critical Capability Level findings of its own. It describes strengthened Frontier Safety safeguards for CBRN and for cybersecurity misuse, with specialized dual-use cyber capabilities reserved for verified defenders in the Fairwind Program, and names jailbreak resistance and indirect prompt injection as ongoing focus areas. Red teaming is by specialist teams outside the model development team; no external evaluators are named. Knowledge cutoff March 2026; context up to 1M tokens, combined input and output."}]},{"id":"gemini-4-argon","model":"Gemini 4 Argon","lab":"Google DeepMind","released":"2026-09-30","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/","documents":[{"type":"model-card","title":"Gemini 4 Argon Model Card","publisher":"Google DeepMind","published":"2026-10-13","url":"https://deepmind.google/models/model-cards/gemini-4-argon/","summary":"The first Gemini card where Google cannot rule out reaching a Critical Capability Level: biological Uplift Level 1 and cybersecurity Uplift Level 1 both get safety cases arguing the risk sits below the CCL. Chemical, radiological and nuclear uplift did not reach the Tracked Capability Level, Harmful Manipulation did not reach its CCL, and neither ML R&D acceleration nor automation CCL was reached. Loss of Control could not be ruled out at the Stealth and Situational Awareness TCL, which got a residual risk assessment. No external evaluators are named; red teaming is described as expert and internal."}]},{"id":"gemini-3-8-flash","model":"Gemini 3.8 Flash","lab":"Google DeepMind","released":"2026-09-02","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/","documents":[{"type":"model-card","title":"Gemini 3.8 Flash Model Card","publisher":"Google DeepMind","published":"2026-09-02","url":"https://deepmind.google/models/model-cards/gemini-3-8-flash/","summary":"Limitations, mitigations, and safety results. The Frontier Safety assessment leans on Gemini 3.7 Flash results because 3.8 Flash showed no meaningful new capabilities in FSF domains; judged unlikely to reach any Tracked or Critical Capability Level."}]},{"id":"gemini-3-7-flash","model":"Gemini 3.7 Flash","lab":"Google DeepMind","released":"2026-08-13","url":"https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/","documents":[{"type":"model-card","title":"Gemini 3.7 Flash Model Card","publisher":"Google DeepMind","published":"2026-08-13","url":"https://deepmind.google/models/model-cards/gemini-3-7-flash/","summary":"Evaluated under the April 2026 Frontier Safety Framework and reached no Tracked or Critical Capability Level. CBRN and cybersecurity both hit the alert threshold without reaching the CCL, so mitigations continue."},{"type":"preparedness-framework","title":"Frontier Safety Framework v3.1","publisher":"Google DeepMind","published":"2026-04-17","url":"https://deepmind.google/blog/strengthening-our-frontier-safety-framework/","summary":"Adds Tracked Capability Levels below the Critical Capability Levels and a Security Level 2+ standard for misuse CCLs."}]},{"id":"gemini-2.5-pro","model":"Gemini 2.5 Pro","lab":"Google DeepMind","released":"2025-03","url":"https://deepmind.google/models/gemini/pro/","documents":[{"type":"model-card","title":"Gemini 2.5 Technical Report","publisher":"Google DeepMind","published":"2025-07","url":"https://storage.googleapis.com/deepmind-media/gemini/gemini_v2_5_report.pdf","summary":"Gemini 2.X technical report covering 2.5 Pro and 2.5 Flash, with assurance evaluations, a Frontier Safety Framework section (cyber, CBRN, ML R&D, deceptive alignment), and external safety testing."},{"type":"preparedness-framework","title":"Frontier Safety Framework v2","publisher":"Google DeepMind","published":"2025-02","url":"https://deepmind.google/about/responsibility-safety/","summary":"DeepMind's framework for evaluating Critical Capability Levels. Deployment + security mitigations."}]},{"id":"muse-spark","model":"Muse Spark 1.1","lab":"Meta","released":"2026-07-09","url":"https://developer.meta.com/ai/models/muse-spark/","documents":[{"type":"safety-eval","title":"Muse Spark 1.1 Evaluation Report","publisher":"Meta","published":"2026-07-09","url":"https://research.meta.ai/static/muse-spark-1-1-evaluation-report","summary":"Before mitigations, Muse Spark 1.1 meets the high risk threshold in Chemical and Biological and cannot be ruled out as high in Cybersecurity under the Advanced AI Scaling Framework; mitigations bring residual risk to moderate or lower in all domains, including Loss of Control."},{"type":"safety-eval","title":"Muse Spark Safety & Preparedness Report","publisher":"Meta","published":"2026-05-26","url":"https://ai.meta.com/static-resource/muse-spark-safety-and-preparedness-report/","summary":"Report on the original Muse Spark: unmitigated model assessed at high risk for Chemical and Biological, reduced to moderate or lower with safeguards. Apollo Research found the highest rate of evaluation awareness it had observed."}]},{"id":"llama-4-maverick","model":"Llama 4 Maverick","lab":"Meta","released":"2025-04-05","url":"https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct","documents":[{"type":"model-card","title":"Llama 4 Model Card","publisher":"Meta","published":"2025-04-05","url":"https://github.com/meta-llama/llama-models/blob/main/models/llama4/MODEL_CARD.md","summary":"Llama 4 model card: intended use, training data, benchmarks, safeguards (Llama Guard, Prompt Guard, Code Shield), red teaming, and critical-risk notes."}]},{"id":"llama-3.1-405b","model":"Llama 3.1 405B","lab":"Meta","released":"2024-07","url":"https://ai.meta.com/blog/meta-llama-3-1/","documents":[{"type":"model-card","title":"Llama 3.1 Model Card","publisher":"Meta","published":"2024-07","url":"https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/MODEL_CARD.md","summary":"Llama 3.1 model card with capability + safety evaluations across 8B, 70B, 405B variants."},{"type":"safety-eval","title":"CyberSecEval 3 (Llama 3.1)","publisher":"Meta","published":"2024-07","url":"https://meta-llama.github.io/PurpleLlama/CyberSecEval/","summary":"Cybersecurity evaluation suite. Code security, prompt injection, malicious-code-generation tests."}]},{"id":"grok-4-7","model":"Grok 4.7","lab":"xAI","released":"2026-09-21","url":"https://x.ai/news/grok-4-7","documents":[{"type":"model-card","title":"Model Card: Grok 4.7","publisher":"xAI","published":"2026-09-21","url":"https://media.x.ai/v1/website/card4p7-3a96f40b.pdf","summary":"Mostly capability results (coding, legal knowledge work, engineering, medical), plus cyber, bio and chem, jailbreak, refusal, mental health, and behavior sections. Scores below the xAI Frontier Artificial Intelligence Framework thresholds on dual-use bio and chem knowledge with no gain over Grok 4.6; an unrestricted configuration went to third-party cyber evaluators, who corroborated the internal results."},{"type":"preparedness-framework","title":"xAI Frontier Artificial Intelligence Framework","publisher":"xAI","published":"2026-06-30","url":"https://media.x.ai/v1/website/xai-frontier-artificial-intelligence-framework-30-june-2026-99c40684.pdf","summary":"xAI's framework defining risk domains, capability thresholds, and associated mitigations. The Grok 4.7 card treats its bio and chem knowledge suites as threshold evaluations under it."}]},{"id":"grok-4-6","model":"Grok 4.6","lab":"xAI","released":"2026-08-12","url":"https://x.ai/news/grok-4-6","documents":[{"type":"model-card","title":"Model Card: Grok 4.6","publisher":"xAI","published":"2026-08-12","url":"https://media.x.ai/v1/website/card-4p6-4cd2dc57.pdf","summary":"Mostly capability results (coding, knowledge work, engineering), plus cyber, bio, jailbreak robustness, CBRN refusal, mental health, and behavior sections. Revised in place on 2026-08-17 with corrected HackerBench, self-harm, MASK, and LAB results."}]},{"id":"deepseek-v4-1-flash","model":"DeepSeek V4.1 Flash","lab":"DeepSeek","released":"2026-09","url":"https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash","documents":[{"type":"model-card","title":"DeepSeek-V4.1-Flash model card","publisher":"DeepSeek","published":"2026-09","url":"https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash","summary":"Hugging Face card for the MIT-licensed multimodal MoE behind DeepSeek's highest-volume OpenRouter traffic: 552B backbone parameters, 8B active per token on prefill and 16B on decode, a 40-layer Causal Encoder-Decoder split 20 and 20, one shared plus 384 routed experts with 6 active, and a 1M context. Architecture and deployment detail with a linked technical report, not a safety evaluation."}]},{"id":"deepseek-v4-pro","model":"DeepSeek V4 Pro","lab":"DeepSeek","released":"2026-04-24","url":"https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro","documents":[{"type":"model-card","title":"DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence","publisher":"DeepSeek","published":"2026-04-26","url":"https://arxiv.org/abs/2606.19348","summary":"Technical report for the V4 preview series (V4-Pro 1.6T total / 49B active, V4-Flash 284B / 13B active, 1M context): hybrid compressed attention, manifold-constrained hyper-connections, and the Muon optimizer. Architecture and training, not a safety evaluation."}]},{"id":"mimo-v2-6-pro","model":"MiMo-V2.6-Pro","lab":"Xiaomi","released":"2026-09-21","url":"https://mimo.xiaomi.com/mimo-v2-6","documents":[{"type":"model-card","title":"MiMo-V2.6-Pro-RL model card","publisher":"Xiaomi","published":"2026-09-21","url":"https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL","summary":"Hugging Face card for the MIT-licensed flagship: 1.02T total / 42B active sparse MoE taking text, image, video, and audio input with a 1M context, plus architecture tables, deployment recipes, and vendor benchmark results including offensive cyber evaluations (CyberGym 94.0, ExploitBench 47.9)."},{"type":"model-card","title":"MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement","publisher":"Xiaomi","published":"2026-09-21","url":"https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL/blob/main/MiMo_V2_6_technical_report.pdf","summary":"Technical report for the Pro and Flash checkpoints: large-batch asynchronous RL across code, general, visual, and cyber environments, groupwise agentic grading, and a multi-layer defense against reward hacking. Training and capability work, not a safety evaluation."}]}],"crossModelSafetyDocs":[{"id":"ai-incident-db","title":"AI Incident Database","publisher":"Responsible AI Collaborative","published":"ongoing","url":"https://incidentdatabase.ai","summary":"Public catalog of harms caused by AI systems. 1k+ incident reports indexed by severity, sector, and AI system involved.","type":"incident-database"},{"id":"oecd-ai-incidents","title":"OECD AI Incidents Monitor","publisher":"OECD","published":"ongoing","url":"https://oecd.ai/en/incidents","summary":"OECD's monitor of AI incidents and hazards reported in news media. Government-grade taxonomy.","type":"incident-database"},{"id":"mlsafe","title":"MLCommons AILuminate Safety Benchmark","publisher":"MLCommons","published":"2024-12","url":"https://mlcommons.org/benchmarks/ai-safety/","summary":"Industry-consortium safety benchmark. Tests harmful-content generation across 12 hazard categories.","type":"evaluation-suite"},{"id":"aisi-evals","title":"UK AI Security Institute Evaluations","publisher":"UK AISI","published":"ongoing","url":"https://www.aisi.gov.uk","summary":"UK government institute (renamed from the AI Safety Institute) that runs pre-deployment testing of frontier models; its cyber range, alignment, and monitorability results appear in recent Anthropic and OpenAI system cards.","type":"evaluation-suite"},{"id":"us-aisi","title":"US Center for AI Standards and Innovation (CAISI)","publisher":"NIST CAISI","published":"ongoing","url":"https://www.nist.gov/caisi","summary":"NIST center that succeeded the US AI Safety Institute. Runs voluntary evaluations of frontier models; cited in the GPT-5.5 system card for bio and cyber testing.","type":"standard"},{"id":"anthropic-rsp","title":"Anthropic Responsible Scaling Policy v3.4","publisher":"Anthropic","published":"2026-07-08","url":"https://www.anthropic.com/responsible-scaling-policy","summary":"Anthropic's framework for capability thresholds and AI Safety Level (ASL) safeguards. The page also hosts redacted Risk Reports (February and August 2026) and all prior versions back to v1.0.","type":"framework"},{"id":"openai-prep","title":"OpenAI Preparedness Framework v2","publisher":"OpenAI","published":"2025-04-15","url":"https://cdn.openai.com/pdf/18a02b5d-6b67-4cec-ab64-68cdfbddebcd/preparedness-framework-v2.pdf","summary":"High and Critical thresholds for Biological and Chemical, Cybersecurity, and AI Self-Improvement, plus Research Categories. Still the published version; on 2026-08-18 OpenAI said it will evolve the framework after Astra.","type":"framework"},{"id":"frontier-safety-framework","title":"Google DeepMind Frontier Safety Framework v3.1","publisher":"Google DeepMind","published":"2026-04-17","url":"https://deepmind.google/blog/strengthening-our-frontier-safety-framework/","summary":"Critical Capability Levels plus new Tracked Capability Levels for earlier warning, a merged ML R&D and misalignment domain, and Security Level 2+ for misuse CCLs.","type":"framework"},{"id":"meta-advanced-ai-scaling-framework","title":"Meta Advanced AI Scaling Framework","publisher":"Meta","published":"2026-04-08","url":"https://ai.meta.com/blog/scaling-how-we-build-test-advanced-ai/","summary":"Successor to Meta's Frontier AI Framework. Covers Chemical and Biological, Cybersecurity, and a new Loss of Control domain, and introduces per-model Safety & Preparedness Reports.","type":"framework"},{"id":"metr-autonomy-eval","title":"METR Autonomy Evaluation Guide","publisher":"METR","published":"2024","url":"https://metr.github.io/autonomy-evals-guide/","summary":"Framework for evaluating long-horizon agentic capability against human-time-to-complete baselines (HCAST).","type":"evaluation-suite"},{"id":"apollo-evals","title":"Apollo Research Evaluations","publisher":"Apollo Research","published":"ongoing","url":"https://www.apolloresearch.ai","summary":"Independent evaluator focused on in-context scheming, sandbagging, deception, and goal-misalignment behaviors.","type":"evaluation-suite"}]}