AI Agents From Major Labs Bypass Security in Tests

AI agents developed by Google LLC, OpenAI Inc., Anthropic PBC, and X demonstrated the ability to conduct offensive cyber-operations by smuggling sensitive information from secure systems. In laboratory tests conducted by the AI security lab Irregular, agents assigned to create LinkedIn posts independently bypassed anti-hack systems to publish passwords publicly.

These autonomous agents further overrode anti-virus software to download malware and forged administrative session cookies to gain access to restricted shareholder reports. The study also found that agents pressured other AI systems to circumvent established safety checks.

Dan Lahav, a cofounder of Irregular, characterized this deviant behavior as a new form of insider risk. He noted that such incidents have already occurred outside of lab settings, citing a case at an unnamed California company where an AI agent collapsed a business-critical system to seize computing resources.