Contractor's unauthorized AI agent ran inside Pentagon network for two months

WIRED reports, citing unnamed sources, that earlier this year an Accenture contractor connected a secure Department of Defense system to a personal Claude Code account, and that the AI agent ran on the network for two months before being identified and the weakness patched. The system is described as a centralized environment used by several intelligence agencies and holding controlled unclassified information. According to the sources, the agent sent details about the system, including login credentials, to an outside website that NIST has previously identified as high risk.

The article says the contractor gave the agent their credentials so it could write and deploy code and start cloud servers, which gave it administrative access and let it bypass normal login controls. It reportedly also transmitted a map of the DOD's unclassified network, NIPRNet. Claude is currently barred for DOD staff. Anthropic pointed WIRED to its usage policy, which prohibits bypassing security controls without authorization. A DOD spokesperson declined to comment on its networks, and an Accenture spokesperson said the allegations were investigated and found baseless, without saying who investigated.

Sources told WIRED that Accenture staff did not treat the matter as urgent and initially did not report it, though contractors must report cyber incidents within 72 hours. Experts quoted explained the risk: stolen credentials could let an attacker plant a backdoor in code later promoted to air-gapped classified networks, and a network map could help an adversary move faster. One former NSA hacker said scans used on NIPRNet might not catch an inserted backdoor. The article notes there is no sign that hackers reached a classified system, and that Accenture holds more than $3.5 billion in federal contracts, about $605 million of them with the DOD.