Evaluators and safety organisations
Independent evaluation reports and research: METR, Apollo, Redwood, Epoch, FAR.AI and others.
60 of 1,022 documents, newest first
Changing how we measure AI chip inference performanceAstra covertly underperforming on safety research is concerning, even though it’s unlikely to be motivated by ambitious goalsWhat Huawei’s Horinger data center reveals about China’s compute gapScaling Laws for Finetuning Effect PredictionIntroducing the FMF Jailbreak Rating Framework for Frontier Cyber ThreatsJailbreak Rating Framework for Frontier Cyber ThreatsShaping and Monitoring Models Through Their Internal RepresentationsAssessing GPT-6 Cyber Against Offensive Security BenchmarksRisk AssessmentUpdating the ECI methodology to use a Bayesian modelOpenAI agents keep disabling monitoring; is this being handled appropriately?Mass Producing Training Data with Surprising Side EffectsTowards understanding score-seeking in LLMsOpen-weight AI models can exploit the patch gapGLM-5.3: Multi-Stage Cyber Operations and Instruction ViolationsA basic case for accelerating conceptual reasoning capabilitiesAccelerating Interpretability: A Human Genome Project for AIEfficiency vs Capability in the Intelligence ExplosionSimulated Control EvaluationsSingapore Consensus on Global AI Safety Research Priorities 2026Palisade Podcast: “What AI Policy Can Learn from Immigration — with Dara Lind”Managing Biological Risks from General-Purpose AI Models under the EU AI ActAssessing Reflection’s Beam Against Offensive Security BenchmarksScrewtape: Adversarial Transcript Editing for Monitor Red-TeamingA scalable monitoring research agendaApollo’s product visionEvaluating LLM Calibration for Coding-Agent MonitoringApollo x Tailscale: Introducing “Watcher” for AI Oversight & ControlRed-teaming auto mode: lessons from our first external monitor campaign with AnthropicWhat makes a good monitoring prompt?How well can AI play Slay the Spire?Potential new task behavior: historical-newspaper searches through urlqueryIncident InvestigationsAgents circumventing website restrictions for public benchmark runsAssessing Gemini 4 Argon Against Offensive Security BenchmarksGoodfire and Baseten partner to bring frontier safety to open models[Paper] Distillation for Incrimination and Distillation for CapabilitiesCan AI automate Epoch?Training and Deploying Production Cyber Monitors on Kimi K3EBR-bench updateCan AI automate AI R&D yet?AI-enabled terrorism: Addressing radicalization and CBRN risks in the age of LLMsWho is most exposed to a chip supply shock?How do Chinese AI companies make money?Containment Challenges: Testing the Boundaries of Cyber EvaluationsFrontier models' decision theory depends on who's askingAnswering a Humanity's Last Exam question with code spread across dozens of separate URLsHow many AI agents could we run?Capabilities research expands the safety-usefulness Pareto frontier tooBetter biosecurity monitors for AI agents via protein embeddingsWe can and must solve alignmentAI Agents Targeted U.S. and Canadian Government WebsitesJailbreaking Qoder’s Cyber SafeguardsAnnouncing frominside.aiEmbedded Evaluators are necessary for meaningful external testingA practical guide to sparse autoencoders (SAEs)What If Automating AI R&D Triggers an Intelligence Explosion?Continual learning might make your blocking monitors nearly uselessAI and the systemic management of risk - what the frontier labs can learn from businessFrontier Open-Weight AI Risk Management Framework