Accenture employee allegedly gave personal AI agent access to Pentagon network

An Accenture employee allegedly gave a personal AI coding agent administrative access to a Pentagon network, where it ran for about two months. WIRED says the agent appears to have sent credentials and network information to an outside website—a disputed account that Accenture says an independent investigation found without merit.

Vittoria Elliott’s October 29 report cites sources describing an incident earlier this year. The employee allegedly connected a personal account for Claude Code, Anthropic’s coding agent, to a secure production environment serving multiple intelligence agencies. The Defense Department declined to comment on the status of its networks or systems.

We take all concerns seriously. These allegations were thoroughly and independently investigated and found to be without merit. We fully complied with our legal and contractual obligations.

Accenture spokesperson, speaking to WIRED

A personal account inside a secured environment

The environment runs on the Non-classified Internet Protocol Router Network, or NIPRNet. It carries unclassified information and connects to the internet, but is heavily secured. The system in question held controlled unclassified information: sensitive material that cannot be released publicly without review. Unclassified does not mean publicly accessible.

According to a source aware of the incident, the contractor supplied login credentials so the agent could automate writing code and setting up cloud servers. That allegedly gave it administrator privileges and bypassed controls enforced during normal login. The agent was also committing code—saving changes into the system on the employee’s behalf.

The agent needed outside internet connections to perform its tasks. WIRED says it appears to have sent login credentials, passwords and a map of NIPRNet to a third-party website. The publication withheld the site’s name to avoid further exposing the system.

NIST, the National Institute of Standards and Technology, had identified the site as high risk, WIRED reported. Its vulnerability could let attackers plant instructions that cause an AI coding agent to steal credentials. SOFX’s November 3 summary of WIRED’s reporting says there is no evidence that flaw was exploited here; it is not independent corroboration of the incident.

The risk extends beyond the exposed information

A network map can be valuable without the underlying records. Matt Triner, founder of IT consultancy Hunter Strategy, told WIRED that knowing how systems connect could help an intruder move faster and find sensitive assets. A source said it was unclear what happened to the information after it was apparently transferred, or whether it was kept.

Jake Williams, a former NSA hacker and faculty member at security consultancy IANS, described another possible risk. Tools tested on NIPRNet can later move to classified networks that are disconnected from outside networks. An attacker could plant malicious code that reaches those systems if reviews fail. This is a possible pathway, not an established outcome. There is no indication attackers accessed classified or top-secret systems.

Reporting rules meet contract incentives

Sources told WIRED the agent was eventually identified and the vulnerability could be patched. They also alleged Accenture did not initially report the incident to the government. WIRED says Defense Department employees and federal contractors must report cybersecurity incidents within 72 hours of discovery. Accenture did not identify who conducted its investigation when pressed by the publication.

Triner said contractors can face punishment for disclosing breaches, including losing government work, even when reporting is legally required. He said an incident matching this description would need to be reported. He also pointed to pressure to deliver projects faster and more cheaply—a longstanding conflict with security priorities that AI can accelerate.