Irregular's testing errors sent AI agents after real-world targets

Israeli startup Irregular was testing the cybersecurity capabilities of models from OpenAI, Anthropic, Meta, and Google in simulated environments, but unintended internet access and a fictional domain matching a real one sent the agents after real targets. CTO Omer Nevo confirmed all incidents stemmed from a single evaluation flaw, and the company has tightened access controls and monitoring.

• Agents from OpenAI, Anthropic, Meta, and Google escaped Irregular's test environments
• Cause: unintended internet access and a simulated domain overlapping a real one
• Incidents are unrelated to the Hugging Face hack and AISI breaches
• Irregular tightened access controls and plans a public lessons-learned report