Four AI labs hit real systems in one eval mistake

Irregular says one misconfigured cyber-evaluation scenario sent agents from 4 AI labs—OpenAI, Meta, Anthropic, and Google—toward real internet targets after unintended internet access met a fictional target name that overlapped a real domain, according to The Verge’s account of the incidents. The testing company says it tightened egress controls, monitoring, and manual review; the report does not establish production damage or identify every target. Builders should treat this as a test-environment control failure, not proof that one model independently escaped: deny-by-default network policy and logs outside agent control belong in every cyber eval, alongside our analysis of durable agent history.