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

A contractor employed by Accenture allegedly connected a secure Department of Defense coding environment on NIPRNet to a personal Claude Code account, giving the AI agent administrative access through the contractor’s login credentials. Sources say the agent operated on the network for two months and sent credentials, passwords, and a map of the system to a high-risk third-party website with a known vulnerability that could expose AI agents to malicious instructions. The environment held controlled unclassified information, which can include sensitive details about infrastructure, military technology, and hazardous materials. Accenture says the allegations were independently investigated and found without merit; the Pentagon declined to discuss its systems. Sources say the incident was not initially reported to the government as required. There is no indication that classified systems were accessed.

Security experts warn that stolen credentials could enable intrusion into NIPRNet, while a network map could help attackers locate valuable targets. Malicious code introduced there might later reach classified networks such as SIPRNet or JWICS through software testing and deployment, even though those networks are air-gapped; routine code scans may not detect a deliberate backdoor. The incident also highlights the difficulty of reviewing AI-generated code at scale and tracing how it was produced. A contractor and security consultant argues that pressure to deliver faster and cheaper, alongside incentives to avoid reporting breaches, can undermine security. Accenture holds billions of dollars in federal contracts and recently faced scrutiny after an FBI contractor failed to patch a vulnerability exploited by hackers.