Anthropic’s METR Access Opens a New Test of Independent Scrutiny
Embedded evaluators may see more, but access limits and publication terms will determine what the public can learn from their findings.
Share
AI Safety & Security
96 posts
New Terrorism Benchmark Puts Modified AI Safeguards Under Scrutiny
New reporting on modified AI models highlights the limits of safety scores, while earlier public research clarifies what counterterrorism evaluations can establish.
Share
Amex’s Purchase-Intent Controls Put Boundaries Around Agent Payments
Public ACE documentation separates customer consent, scoped credentials, and planned protections, with important limits on what can be verified today.
Share
AISI Wants Frontier AI Oversight to Begin Before Public Release
AISI's new technical agenda extends oversight across development and deployment while distinguishing voluntary best practice from mandatory requirements.
Share
Kyvvu’s Agent Security Pitch Depends on the Whole Task, Not One Permission
Kyvvu's action-control approach follows task history, while its documentation shows why policy design and complete integration remain central to the security claim.
Share
OpenAI Researchers’ Dismissal Dispute Puts Independent Safety Access in Focus
Three former OpenAI researchers contest misconduct allegations and call for external access, monitorability and clearer rules for safety collaboration.
Share
Anthropic’s Offline Evaluation Shift Leaves a Restart Question
Anthropic's reported restriction on internal evaluations prompts an editorial assessment of external exposure, agent behaviour and the evidence behind a future restart.
Share
Teen Chatbot Safety Study Questions Engagement Cues as OpenAI Disputes Its Methods
Common Sense Media reports persistent relational and engagement signals in ChatGPT for Teens. OpenAI challenges the test setup and offers separate usage figures.
Share
Meta Targets Hidden Abuse Links as Its Enforcement Figures Leave Detection Limits Unclear
Meta reports action against 33.2 million content items and describes new destination checks and internal testing. The figures do not measure all undetected abuse.
Share
Malaysia Links Its AI Nation Ambition to Data Security and Digital Trust at CyberDSA
CyberDSA's opening puts secure infrastructure, data access and trust alongside Malaysia's 2030 AI ambition. Exhibits and planned agreements do not establish deployment readiness.
Share