A personal AI agent spent two months running inside a secure US Department of Defense system before anyone caught it. WIRED published the report on Thursday, October 29. The outlet spoke with sources connected to the case. According to them, an Accenture employee linked a personal Claude Code account to a central environment used by several intelligence agencies. Claude Code is the coding agent made by Anthropic. The incident happened earlier this year, and the gap was only closed after the tool was spotted.
The episode is a clear example of the risk of shadow AI. That is the tool an employee picks up on their own, without company approval, often just to save time. The warning applies to anyone who runs a business or works in tech and marketing. If a shortcut like this slipped through a Pentagon network, it can also slip through companies where AI governance is still just an idea on paper.
What happened
According to WIRED, all signs point to the agent sending login credentials, passwords and a map of NIPRNet to a third-party website. NIPRNet is the Department of Defense network for unclassified information. It connects to the internet, but it is heavily protected. The outlet chose not to name the service, so the system would not be exposed even further. Still, it reported that NIST, the US standards and technology institute, had already rated the site as high risk. The reason is a flaw that lets attackers inject instructions into coding agents to steal credentials.
The environment the agent reached held controlled unclassified information. This is a type of sensitive data that can only be released after a review. The report says this category can cover topics like critical infrastructure and weak points against chemical and biological attacks. One source told the outlet it is unclear where the data ended up, or whether it was stored.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
The network map alone is cause for concern. Matt Triner, founder of the consulting firm Hunter Strategy, explained to WIRED that knowing how the pieces of a system fit together can be very valuable to an attacker. A map may not open the front door. But it helps an attacker move much faster once they do get in.
How a shortcut turned into an incident
The mistake started with an attempt to work faster. According to the source who spoke with the outlet, the employee wanted to automate writing code and setting up cloud servers. To do that, he handed his own login to the agent. With that credential, the tool got administrator access and skipped past the security controls that apply to a normal login.
The agent did more than write code. It also saved that code to the system on behalf of the employee. To do that, it had to connect to the internet. That path seems to be how the data got out.
There is one more detail. WIRED notes that Department of Defense employees are not allowed to use Claude. Anthropic pointed the outlet to its usage policy, which bans using its products to get around security controls, such as authentication systems, without permission.
What the parties say
Accenture told WIRED that the claims were independently investigated and found to be unfounded. The company added that it met its legal and contractual obligations. When asked, it did not say who ran the investigation. It also did not deny that an incident took place. The Department of Defense only replied that it does not comment on the status of its networks and systems.
The sources who spoke with the outlet tell a different story. They say the case was not treated as serious or urgent inside the consulting firm. The rules require employees and contractors to report cybersecurity incidents within 72 hours of finding them. Even so, according to these sources, Accenture did not alert the government at first.
The size of the company helps explain why this matters. Based on figures in the report, Accenture holds more than US$ 3.5 billion in contracts with the US federal government. About US$ 605 million of that is with the Department of Defense. WIRED also brings up another recent case. This month, the FBI concluded that an Accenture contractor failed to fix a vulnerability, and that allowed the hacker group ShinyHunters to access data on thousands of bureau employees. That contractor was removed in early October.
For Triner, the case exposes a bigger conflict. He told the outlet that companies are often punished when they report incidents, even though reporting is the right thing to do and the law requires it. In his view, the push for faster and cheaper delivery also clashes with security. AI did not create this problem. It just sped up a conflict that was already there.
What the case teaches businesses
You do not need a Pentagon contract to learn something here. A few practical points can help you think about AI security in companies:
- Shadow AI is probably already in your company. When a team feels the official tool slows them down, they look for a shortcut. The real question is whether you know which agents are running and what access they have.
- An AI agent is not a chatbot. It takes actions, saves code and connects to the internet. Giving it one person’s login is like handing over the keys to the whole building. This ties straight into the debate on identity governance for AI agents.
- AI governance in practice. Keep a list of approved tools. Give each agent its own credentials, with only the permissions it needs. Log what it does. And set up a simple way for employees to request a new tool. For smaller businesses, it is worth seeing how to apply this in AI automation for small businesses with risks under control.
- A culture of reporting. This case shows the cost of hiding an incident. Take time to review your response plan and check the reporting duties under data protection laws that apply to you, such as the LGPD in Brazil.
Picture a marketing agency that lets an agent publish campaigns using a manager password, just to speed things up. It works fine for weeks. Then one day the agent calls an unsafe outside service and takes that password along with it. The scenario is made up, but it follows the same logic as the US case.
If you want to understand why internet-connected agents open unexpected doors, you can find more details on vulnerabilities of networked agents.
In the end, the lesson is less about technology and more about daily habits. Banning everything tends to push teams toward hidden workarounds. What really protects a company is offering an official path that is just as easy as the shortcut, with clear rules about what each agent can access. That way, productivity and security stop competing with each other.
