A cybersecurity startup claims its small language model, post-trained with inline reference monitoring, outperforms GPT5.5-xhigh on coding agent security benchmarks including LinuxArena and SleightBen…
A new methodology using ELO ratings to measure red team and blue team capability in AI control research has been proposed, aiming to answer whether monitors improve faster than attackers as models bec…
Sandboxing increases safety against untrusted AI coding agents, according to a study on LinuxArena that tested ten sandboxing protocols against an untrusted agent (Sonnet 4.5) and a trusted monitor (G…
AI control research must expand from models to agent harnesses as frontier labs adopt harnesses with skills, memory, subagents, and external services by 2026, according to a LessWrong analysis. Claude…
Researchers developed a method to combine AI control protocols by routing between them based on predicted usefulness and suspicion, achieving better safety-usefulness tradeoffs. They also addressed at…