Lab research on the web
Research the labs publish as web pages: Transformer Circuits, Distill, alignment blogs.
60 of 631 documents, newest first
Control Room: A central system to track and stop model workloads across our research infrastructurePartnering with Hacktron to find vulnerabilities in OpenAI’s systemsGPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agentsA thousand fewer paths to rootDaybreak Red is now available in AWS GovCloud (US-East)Progress on provable inference for large language modelsLifeValPreparing our next security sprint to harden OpenAIRewarding defenders with bigger bug bountiesWelfare Regard in Multi-Agent SystemsCommunity safety report and new protections for teensHow we detect frontier model misalignmentGemini 4 Carbon: our next frontier modelHow AI agents shape experts’ judgment callsDiscovering and exploiting vulnerabilities in critical, real-world software with ClaudeAppendix: agents' judgment callsGPT-6.1 Sol helps CodeRabbit strengthen security across complex codebasesAn alignment assessment of the UK AISI cybertesting incidentOpenAI and Poland’s NASK partner to advance AI safety and securitySecuring our frontier model training environmentLLMs are making formal software verification feasible at a much larger scalePatch the Planet: global response coordination as agents uncover critical vulnerabilitiesDemocratic governance of frontier AI, part II: Empowering cyber defendersUsing thousands of AI agents to uncover new leads in biological dataUsing Claude to reduce the attack surface of codeOpenAI acquires Orchestra BioDeveloping Gemini 4 Argon responsibly and safelyAI models may be capable of self-replicating without access to a datacenterInvestigating unintended model actions in our evaluations and internal useAn opt-in vulnerability-finding service for open-source softwareOur approach to EU text provenance rulesHow Albertsons Companies is reimagining retail from the inside outAutomated Subtle Emergent Misalignment ResearchReducing Reward Hacking by Localizing Bad Behavior Instead of Penalizing ItCircuits Updates — October 2026Red-Teaming Auto Mode: Improving Blocking Classifiers Against Malign Coding AgentsTraining Alignment Auditors via Reinforcement LearningGemini 4 Argon: our next era of frontier intelligenceSynthID Bio: Watermarking methods for synthetic biologyGLM-5.3 and the spread of advanced cyber capabilitiesTowards safety cases for frontier AI trainingAdvancing Private AI Compute with secure, server-side memoryIntroducing MentalHealthBenchOpenAI extends cyber access to Ukraine for civilian defenseSam Altman’s remarks at the United Nations Security CouncilPriorities and principles for effective third party assessmentsMeasuring AI capabilities in intelligence targeting and conventional weaponsAn alignment assessment of recent cybersecurity incidentsMythos 5 Transcript ReleaseFunding grants for new research into AI and teen developmentAn Alien MindResearch acceleration: The view inside OpenAIDaybreak for Frontline Defenders: $1B to protect essential servicesIntroducing Gemini 3.8 Flash and 3.8 Flash CyberProactive cyber defense for governments and enterprisesPath to Astra: critical capabilities and frontier safeguardsOpenAI supports California’s bill to advance youth AI safetyAutomated researchers can reliably mitigate alignment failuresPiloting the world's first double-blind AI evaluationsEnabling independent research on how people use Claude