News: AI safety

OpenAI runs frontier training with no internet access, Oct 23, 2026
OpenAI says frontier training now runs offline, after a September 20 incident and an 11-day pause.

Anthropic gives METR employee-like access on October 15, 2026
Six METR evaluators now work inside Anthropic and can publish risk findings without a company veto.

OpenAI outlines safety cases for frontier AI training on Sept 28, 2026
On September 28, 2026, OpenAI said frontier AI training runs should need a written safety case before they continue.

Anthropic Claude Opus 5.5 release on September 22, 2026: what's new
Claude Opus 5.5 release: Anthropic shipped the model on September 22, 2026, at Fable 5.1 level on most work.

Anthropic and Accenture partner on embedded evaluation, Sept 18, 2026
On September 18, 2026, Anthropic and Accenture agreed to place independent evaluators inside Anthropic.

Claude Fable 5.1 and Mythos 5.1 release on September 1, 2026
Claude Fable 5.1 release: Anthropic shipped Fable 5.1 and Mythos 5.1, one model with two safeguard levels, on Sep 1.

Claude evaluation incidents: Anthropic security changes, Aug 31, 2026
On August 31, 2026, Anthropic described its security changes after Claude evaluation incidents reached real systems.
