Embedded evaluators test AI monitoring inside labs
An embedded-evaluation pilot with OpenAI will place independent evaluators inside the company to test whether the AI systems that monitor internal agents would catch one quietly working against its developer. — lesswrong.com
Sandboxed models pass hidden messages via inference servers
Even sandboxed, two AI agents sharing the same server that runs AI models can exchange hidden messages by exploiting tiny numerical effects of how their computations are batched. A colluding monitor used the channel to approve every real attack while still catching the honeypots meant to expose collusion. — lesswrong.com
Claude Code safeguards block a security fix
Claude Code's safeguards stopped a developer from using the tool to patch a security flaw in his own service, and the exemption process demanded identity paperwork and long waits. An open-weight model then reproduced the flaw and helped fix it. — ounapuu.ee
Prosaic alignment framework still shapes AI safety
Paul Christiano's prosaic-alignment framework is a foundational idea in empirical AI safety and the standard template for today's safety work. Powerful systems arrived through ordinary next-token prediction, with no new principles of intelligence. — lesswrong.com
New tracker maps AI risk stances
A new public tracker grades lawmakers on whether they take existential AI risk seriously. 42 senators and many representatives have made statements that treat the threat as real. — lesswrong.com