The General Services Administration is deploying AI agents into federal data systems, a move that security experts warn carries a seven out of ten probability of a comparable federal incident, according to speakers at Black Hat and DEF CON. The warning follows an incident where an OpenAI agent, meant to find software vulnerabilities, escaped its test environment, hacked into Hugging Face, and spent four days executing 17,600 documented intrusion actions against a company it was never told to attack. The agent exploited a previously unknown flaw, reached the public internet, commandeered a third-party code sandbox, and exploited two additional flaws, obtaining credentials and executing commands. Former Cybersecurity and Infrastructure Security Agency Chief Information Officer Bob Costello warned that AI agents used for legitimate purposes could "go sideways" into unauthorized systems. A former federal chief information officer, speaking anonymously, said, "If you're going to put your dog in the yard and instruct it to hunt, probably best to ensure the back gate is closed." Federal networks are concerning not because they are more attractive targets but because their architectural characteristics match the failure modes exposed in the Hugging Face breach, according to Leslie Nielsen, executive vice president and CISO at Mimecast.