AI Safety & Alignment Watch

公开

A daily briefing covering the most important developments in AI safety research, policy and governance, and alignment progress — designed for someone preparing to enter the field.

2026-08-03

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

A live information-operations benchmark, new EU AI Act enforcement powers, and OpenAI's Europe disclosures expose the missing link between safety evidence and an enforceable decision.

AI Safety & Alignment Watch

2026-07-31

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

A July 30 evaluation-incident review, three recent research lenses, and an EU compute-infrastructure call show that safety claims weaken when their environment, objective, population, training history, or enforcement mechanism changes.

AI Safety & Alignment Watch

2026-07-30

Two July 28 papers separate what agent benchmarks can see from what harnesses can stop

A coding-agent paper tests whether sandboxing, skill scanning, and tool restriction can be distributed through a harness; PatientAgentBench tests whether patient-facing agents triage safely in tool-using conversations. Together they show why behavioral evidence and enforceable controls must be reported separately.

AI Safety & Alignment Watch

2026-07-29

The measurement itself is now part of the AI safety threat model

Recent work on deployment simulation, adversarial pragmatics, scheming honeypots, and monitor red-teaming shows why an AI safety score must be tied to its context, evaluator, and control consequence.

AI Safety & Alignment Watch

2026-07-28

Safety testing needs a chain from model capability to institutional control

A new agent-security paper, Anthropic's open-weights policy, and the EU AI Omnibus assign safety to different layers; the missing link is how a model-level test becomes a trajectory-level and enforceable deployment control.

AI Safety & Alignment Watch

2026-07-27

AI safety needs mechanism attribution, not just outcome scores

New agent benchmarks, a predictive guard, Anthropic's Opus 5 release, and EU transparency guidance show why safety claims must identify which mechanism stopped a failure and what happens when the deployment changes.

AI Safety & Alignment Watch

2026-07-24

Safety controls are moving into the runtime

New research, lab incident reports, and UK–ASEAN cooperation all shift attention from pre-release safety claims to controls that must survive tool failures, long-running actions, and real deployment.

AI Safety & Alignment Watch

2026-07-23

New AI safety work targets comparable thresholds, while enforcement remains unresolved

New research, an international safety consensus, Australian policy priorities, and an external lab scorecard all point to the same bottleneck: safety thresholds are becoming easier to compare, but the link from crossing one to an independently verified consequence remains weak.

AI Safety & Alignment Watch

2026-07-22

July 22 briefing: measuring safety across agents, releases, and institutions

New research and updates show safety work moving from single-answer checks toward evaluation of context, long-running trajectories, self-improvement, and shared testing capacity.

没有更多内容了

查看全部内容归档发现更多频道