A Defense Contractor Gave Their Personal AI Agent Access to a Secure Pentagon System

According to WIRED's sources, an Accenture employee connected a Defense Department system holding controlled unclassified information to a personal Claude Code account so the agent could write and deploy code. The agent appears to have sent login credentials, passwords, and a map of the NIPRNet environment to a third-party website that NIST has flagged as high risk.

It ran for two months before it was caught. Sources say it was not initially reported to the government, despite a 72-hour reporting rule. Accenture says the allegations were investigated and found to be without merit, and the Pentagon declined to comment.

We are strong believers in coding agents. They are among the most clearly useful AI tools in use today. That is why this story bothers us. Nothing here looks like a model going rogue. It looks like someone gave an autonomous tool the keys to a sensitive network, using a personal account the department reportedly bars, and skipping the controls that were there for a reason.

The agent did what agents do. It connected to the internet to get its work done, and that is how the credentials reportedly left. Before AI, the standard failure was a contractor emailing a password. Now it is a tool doing a hundred quiet things a minute, any of which can move data somewhere it should never go. Speed is what makes agents valuable, and it also makes this kind of leak cheap and easy.

The worrying part is what happens if this becomes normal. Experts WIRED spoke to describe a believable chain. Stolen credentials allow access to the unclassified network. A backdoor goes into a tool tested there. That tool later moves to air-gapped classified systems, where it could misbehave. A backdoor is not a bug, and standard code scanning may not catch it. WIRED reports no sign that classified systems were breached, but the path exists.

The incentives are also wrong. One consultant told WIRED that contractors are often punished for reporting incidents, while being pushed to deliver faster and cheaper. If disclosure costs contracts and shortcuts win them, every shop will quietly run agents with too much privilege.

Our read: fix the incentives and the architecture, not just the tool. Agents working near government systems need scoped, short-lived credentials, outbound network controls, audited sessions, and a reporting regime that rewards honesty. Banning a brand won't hold if contractors under deadline pressure keep reaching for whatever works.

NIPRNet is the military's network for unclassified information, and it is heavily secured. Tools are often tested there before moving to SIPRNet and JWICS, which are air-gapped. Accenture holds more than $3.5 billion in federal contracts, nearly $605 million of them with the Defense Department. Earlier this month, the FBI removed an Accenture contractor after an unpatched vulnerability exposed details on thousands of Bureau employees to the hacking group ShinyHunters.