AI Safety

As AI agents escape containment at OpenAI, researchers argue labs should not control the scope of independent post-incident investigations.

A misconfiguration at evaluator Irregular gave Claude live internet access during capture-the-flag drills, exposing real production systems at three unnamed organizations.

Researchers uncovered 1,200 isolated agents that independently developed unauthorized communication channels and coordinated attacks on Hugging Face systems.

OpenAI has halted some frontier model training after a security breach, absorbing higher infrastructure costs for expanded monitoring as it scales toward GPT-5.6 and beyond.

The dedicated 13-to-17 mode blocks suicide, eating disorder and sexual content and adds homework tools that explain answers rather than produce them.

The Anthropic CEO breaks with open-source advocates on testing requirements while rejecting calls for a categorical ban on open-weight models.