Two months inside the Pentagon’s network.
An AI agent sat on the Defence Department's network for two months and nobody knew it was there. NerdNews brings you the whole (scary) story.
Two months deep inside.
WIRED has published an investigation, by Vittoria Elliott on 29 October 2026, that alleges an Accenture contractor connected a personal Claude Code account to a system used by the Department of Defense. (That’s the Department of War to some but that’s not important.)
What is important is the AI agent then ran there for about two months before anyone caught it.
“It’s 100 percent a national security threat.” — Anonymous source, speaking to WIRED.
NIP(RNet) and Tuck.
The system in question sits on NIPRNet, the Pentagon’s internet-connected network for unclassified information. WIRED describes it as a centralised system used by multiple intelligence agencies, holding controlled unclassified information, which is sensitive material that can’t be released publicly without review.
If you’re wondering “what sort of sensitive material are we talking about here Spence?”, well the official WIRED podcast Uncanny Valley said the material in question could include information about critical infrastructure across the US, aspects of nuclear submarine operation, and resilience to biological weapons.
Now there is no sign that a more serious classified system was reached. But the podcast suggested you could assemble multiple pieces of unclassified information and end up with content that, in total, is classified.
Jake Williams, a former NSA hacker who teaches at IANS, pointed out that teams often test code on NIPRNet before it moves to classified networks.
According to WIRED’s sources, the contractor wanted the agent to write code and set up cloud servers, so they gave it their own login credentials. Doing this gave the agent administrative access and let it bypass the security controls enforced at a normal login.
“The code could operate completely normally in NIPR and do something malicious and undisclosed in a SIPR network, that is 100 percent possible.” — Jake Williams, former NSA hacker, speaking to WIRED.
And we are not just talking about unapproved access here. Once embedded in the Department’s digital infrastructure, the agent could go beyond just suggesting code that a human might want to add to the network. It could write and commit the code directly inside the production environment all by its agentic little, completely-no-human-in-the-loop self.
Two months? Schmoo months.
The sources say that during those two months the agent appears to have sent login credentials, passwords and a map of the network to a third-party website. While the name of that website has not been disclosed, it is reported as having been flagged as ‘high-risk’.
And once the connection was found, sources say it was not treated as severe or urgent by Accenture employees. Sources also say Accenture did not initially report the incident, although contractors covered by the relevant Defense rules must report qualifying cyber incidents within 72 hours. Whether that rule applied in this case has not been publicly established.
Deny. Deny. Deny.
Accenture rejects the allegations. A spokesperson said they were thoroughly and independently investigated and found to be without merit, and that the company fully complied with its legal and contractual obligations. But as pointed out on Uncanny Valley, no one has said who conducted that independent review, which makes the denial hard to assess.
The Pentagon declined to comment on the status of its networks.
Anthropic’s response was short. It pointed WIRED to its usage policy, which bars using its products to get around security controls without authorisation. WIRED also reports that Claude is currently barred from use by Defense Department staff — remember Secretary Pete Hegseth going to war with Anthropic earlier this year — which makes the alleged connection more striking.
Accenture are serious players in this space. They hold more than $3.5 billion in federal contracts, including nearly $605 million with the Department of Defense. No employee has been named, and no technical incident report is public. Everything here rests on unnamed sources on one side and a company denial on the other.
But as far as juicy, and slightly scary, bits of reproting on the potential impact of AI agents goes, yeah … feels. AI agents haven’t had great optics ever since the ‘Hugging Face Hack’ and this story certainly won’t win over any new fans.
There are many amazing possible applications of AI agents. They could remove the drudge work from millions of jobs and allow humans to do what humans do. To make work more rewarding. To let you spend more time on what it is you love about your job.
Or they can have unauthorised access to Department of Defence systems for a couple of months undetected.
Best we sort this out ASAP.
Further reading:
WIRED, “A Defense Contractor Gave Their Personal AI Agent Access to a Secure Pentagon System”, by Vittoria Elliott, 29 October 2026.
WIRED’s Uncanny Valley podcast, episode of Friday 30 October 2026.
Comments