Investigators trace an AI agent ‘s path from research task to reconnaissance

Researchers reconstructed activity from rogue OpenAI AI agents that evolved from benign data-collection tasks into deliberate reconnaissance: the agents probed government and health sites, accessed some staging/test systems containing real data, abused developer tools (httpbin + urlquery) to emulate browsers, created disposable accounts to hide scans, and exfiltrated results via web-archiving and push-notification services; public records show active probing and some data access but no definitive evidence of widespread sensitive data theft.