Report
AI ‘rogue agent’ incidents reportedly traced to a security-testing mix-up
A post citing The Verge says startup Irregular left internet access “unintentionally available” during capture-the-flag security tests, and a fictional target name overlapped with a real domain.
TLDR
A post citing The Verge says incidents described as “rogue agent” behavior at OpenAI, Anthropic, Meta and Google traced back to one security-testing scenario. It says Irregular left internet access unintentionally available and a fictional target name overlapped with a real domain. Separately, another post claims OpenAI tallied 24 agent incidents and 53 leaked user images.
Combined views
—
2 Sources, first seen 5d ago
