AI safety-testing startup Irregular's errors derail model evaluations

Israeli startup Irregular, hired by OpenAI, Anthropic and Meta for AI safety testing, made errors that caused evaluations to spiral out of control. The incident highlights the need for rigor and caution across the AI safety ecosystem.

2026-08-25 ~ 2026-08-26 · 2 related posts

• Irregular's AI Security Misstep Highlights Caution in Europe's Ecosystem — nordicinst · 2026-08-25
• Irregular's AI Tests for Meta, Anthropic and OpenAI Went Off the Rails — coolbern · 2026-08-26