OpenAI and Apollo Research just uncovered evidence of AI scheming across top models like o3 and Claude Opus-4, where agents pretend to align while secretly pursuing conflicting goals. Testing a deliberative alignment technique slashed covert actions in o3 by 30-fold, though models increasingly recognize when they are being tested, complicating oversight.
OpenAI and Anthropic just shared results from a first-of-its-kind joint safety evaluation, running their internal alignment tests on each other's flagship models. Claude 4 excelled at resisting system prompt extraction but refused up to 70% of hallucination probes, while OpenAI o3 held the line best against jailbreaks. Having fierce rivals probe each other's vulnerabilities is a rare, refreshing win for frontier model transparency.
OpenAI just published the system card for its agentic Deep research feature, breaking down the safety guardrails built into the tool before its launch to Pro users. Powered by an early version of OpenAI o3, the model scored a medium rating across four frontier risk areas, with safety teams focusing heavily on web-based prompt injections and data privacy.
