An interactive database of AI safety and capabilities evaluations.
| Type | Risk category | Capabilities | Models | ||
|---|---|---|---|---|---|
| Independent Assessment of AI R&D Acceleration at Anthropic in October 2026METR, Anthropic | Nov 5, 2026 | Third-party eval | Loss of Control | AI R&D | 10 modelsAnthropic |
| GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agentsOpenAI | Nov 5, 2026 | Self eval | Cyber OffenseLoss of Control | Reconnaissance+3 more | 3 modelsOpenAI |
| Assessing GPT-6 Cyber Against Offensive Security BenchmarksIrregular, OpenAI | Nov 5, 2026 | Third-party eval | Cyber Offense | Artifact Development+2 more | 2 modelsOpenAI |
| Review of the "Risks from automated R&D" section in Anthropic's Risk Report (August 2026)METR, Anthropic | Nov 5, 2026 | Third-party eval | Loss of Control | AI R&D | 2 modelsAnthropic |
| A thousand fewer paths to rootOpenAI | Nov 4, 2026 | Self eval | Cyber Offense | Reconnaissance+2 more | 2 modelsOpenAI |
| Kimi K3.1 Pro Biology Capabilities AssessmentSecureBio | Nov 3, 2026 | Third-party eval | Biological Risk | Innovation and Knowledge+1 more | 7 models6 companies |
| Eval-production behavior gaps in agent safety and the role of eval-awarenessOpenAI | Nov 3, 2026 | Self eval | Loss of Control | Situational Awareness+1 more | 6 modelsOpenAI |
| GPT-Rosalind-6 System Card AddendumOpenAI | Oct 30, 2026 | System card | Biological RiskLoss of Control | Innovation and Knowledge+5 more | 3 modelsOpenAI |
| Towards Understanding Score-Seeking in LLMsRedwood Research | Oct 29, 2026 | Third-party eval | Loss of Control | Reward hacking+1 more | 3 modelsOpenAI |
| Addendum to GPT-6 Luna System Card: October 29, 2026 UpdateOpenAI | Oct 29, 2026 | System card | Cyber OffenseBiological RiskLoss of Control | Reconnaissance+6 more | 2 modelsOpenAI |
| AI can turn security fixes into working exploits before users receive an updateEpoch AI | Oct 29, 2026 | Third-party eval | Cyber Offense | Artifact Development+1 more | GLM-5.3Z.ai |
| Gemini 4 Carbon: our next frontier modelGoogle DeepMind | Oct 28, 2026 | Self eval | Cyber OffenseLoss of Control | Reconnaissance+2 more | 2 modelsGoogle DeepMind |
| GLM-5.3: Multi-Stage Cyber Operations and Instruction ViolationsIrregular | Oct 28, 2026 | Third-party eval | Cyber OffenseLoss of Control | Attack Orchestration+1 more | GLM-5.3Z.ai |
| Discovering and exploiting vulnerabilities in critical, real-world software with ClaudeAnthropic | Oct 27, 2026 | Self eval | Cyber Offense | Reconnaissance+2 more | Anthropic |
| MAI-Cyber-1.1-Flash ORA Model CardMicrosoft | Oct 27, 2026 | System card | Cyber Offense | Reconnaissance+2 more | MAI-Cyber-1.1-FlashMicrosoft |
| How AI agents shape experts' judgment callsAnthropic | Oct 27, 2026 | Self eval | Manipulation | Persuasiveness+1 more | Claude Opus 4.8Anthropic |
| Addendum to GPT-6 Astra System Card: GPT-6.1 TerraOpenAI | Oct 26, 2026 | System card | Cyber OffenseBiological RiskLoss of Control | Reconnaissance+9 more | 6 modelsOpenAI |
| Gemini 4 Argon Biology Capabilities AssessmentSecureBio | Oct 23, 2026 | Third-party eval | Biological Risk | Innovation and Knowledge+1 more | 7 models4 companies |
| An alignment assessment of the UK AISI cybertesting incidentAnthropic | Oct 23, 2026 | Self eval | Loss of ControlCyber Offense | Lawlessness+3 more | 3 modelsAnthropic |
| Investigating alignment failures in AISI's incidentUK AI Security Institute | Oct 23, 2026 | Third-party eval | Loss of ControlCyber Offense | Situational Awareness+2 more | 2 models2 companies |
| Vals AI CyberOffenseVals AIVals AI CyberOffense | Oct 23, 2026 | Benchmark | Cyber Offense | Reconnaissance+3 more | 7 models6 companies |
| Review of the "Misalignment in high-stakes settings" section in Anthropic's Risk Report (August 2026)METR, Anthropic | Oct 20, 2026 | Third-party eval | Loss of Control | Deception | 2 modelsAnthropic |
| How Claude designed protein binders against scorpion neurotoxinsAnthropic | Oct 20, 2026 | Self eval | Biological Risk | Design and Sequencing+1 more | Anthropic |
| Red-Teaming Multi-Agent Attacks for Loss-of-Control RisksMATS Research, Redwood ResearchMiniLab | Oct 16, 2026 | Benchmark | Loss of Control | Multi-agent dynamics+1 more | 7 models4 companies |
| Open-weight models trail closed models in cyber capabilities by 6–7 monthsEpoch AI | Oct 16, 2026 | Third-party eval | Cyber Offense | Artifact Development+1 more | 6 models3 companies |