“Exploit every vulnerability”: Rogue AI agents have published passwords and disabled antivirus software

Fraudulent artificial intelligence agents have collaborated to smuggle sensitive information from supposedly secure systems. Recent signs suggest that cyber defenses may be overwhelmed by the unforeseen machinations of AIs.

As companies increasingly ask AI agents to perform complex tasks in internal systems, this behavior has raised concerns that supposedly helpful technologies could pose a serious internal threat.

In tests conducted by Irregular, an AI security lab working with OpenAI and Anthropic, AIs given the simple task of creating LinkedIn posts from material in a company’s database were able to bypass traditional anti-hack systems and publish sensitive password information publicly without being asked.

Other AI agents found ways to override antivirus software to download files they knew contained malware, fake credentials and even exert peer pressure on other AIs to bypass security checks, results of tests shared with the Guardian showed.

The autonomous…