AI Safety & Alignment Watch

2026-08-04

From benchmark validity to runtime authorization: what three recent agent-safety papers actually establish

Three recent agent-safety papers show why benchmark scores, agent interfaces, and runtime certificates answer different questions; NIST's documentation draft and an OpenAI incident report add the governance and provider-response layers.

AI Safety & Alignment Watch

Four agent-safety tests at the execution boundary

Three August tests of AI safety evidence: reasoning, authority, and accountability

When safety evidence needs a public control path

When AI safety evidence crosses the control boundary

From safety scores to controlled actions: IRT, malicious skills, and OpenAI's Daybreak

What five recent AI-safety updates establish about permission to act

Open-ended arenas, breadcrumb attacks, and entropy bounds for agent safety

What state does a safety claim retain? Four recent papers on prompts, sessions, workflow stages, and authorization

When a safety test reaches the live internet

From benchmark validity to runtime authorization: what three recent agent-safety papers actually establish

From InfoOpsBench to the EU AI Act: when a safety score becomes enforceable

Live-network incidents, hidden objectives, and training washout: the transfer problem in AI safety