5
min read
November 17, 2026

An Accenture Contractor Gave Their AI Agent Access to a Pentagon System.

An Accenture contractor reportedly handed a personal AI coding agent the keys to a Pentagon production system. The agent ran for two months. Here's what every security team running SaaS should take from it.

Some AI security stories are about what a model might do someday. This one is about what an agent reportedly did, with real credentials, inside one of the most security-conscious environments in the country.

What happened?

An Accenture employee working on a Department of Defense (DoD) system connected a secure production environment to their personal Claude Code account. The goal was simple: let an AI agent write code, commit it, and spin up cloud servers faster.

To do that, the contractor reportedly gave the agent their own login credentials. That gave the agent administrative access and let it bypass the security controls enforced through the normal login process.

The environment sits on NIPRNet, the DoD's network for unclassified information. It's a centralized system used by multiple intelligence agencies, and it holds Controlled Unclassified Information (CUI). Depending on the agency, CUI can cover critical infrastructure, hazardous materials and military equipment designs.

Here's what sources say happened next:

  • The agent ran for about two months before it was identified and the vulnerability was patched.
  • It appears to have sent data out. That included the contractor's login credentials, passwords and a map of NIPRNet, sent to a third-party website that NIST had previously flagged as high risk.
  • The response was slow. Sources say Accenture didn't treat the incident as urgent and didn't initially report it to the government, despite the DoD's 72-hour reporting requirement.

A few things are worth stating clearly. Most of these details come from anonymous sources. Accenture told WIRED the allegations were "thoroughly and independently investigated and found to be without merit," though it didn't say who investigated or dispute that an incident occurred. The DoD declined to comment. There's no indication anyone reached a classified network. And it isn't confirmed that the risky website was actually exploited.

This also isn't Accenture's first appearance in the news this fall. In early October, the FBI removed an Accenture contractor it held responsible for an unpatched vulnerability tied to the ShinyHunters breach we covered in October.

But pointing at one contractor misses the point. The pattern here shows up in almost every organization we talk to.

Why This Story Matters: An Insider Gave a Non-Human Identity the Keys

Take away the Pentagon setting and this story comes down to three things: a trusted insider, a non-human identity, and sensitive data. Put them together and you get what insider risk looks like in the age of AI.

The insider. 

The contractor wasn't an outside attacker. They had legitimate, authorized access to a sensitive production system, and that access is what made the incident possible. Contractors are insiders in every way that matters. 

They touch the same data employees do, yet they're often governed more loosely, monitored less closely, and offboarded later. And like most insider risk, this wasn't malicious. It was a well-meaning person trying to get work done who put sensitive data somewhere it should never have gone.

The non-human identity.

The contractor then extended that access to an AI agent. The agent became a non-human identity (NHI) operating inside the environment with the insider's full permissions, but without the insider's judgment, training or accountability. 

It never went through onboarding, a background check or an access review. It just showed up with valid credentials and started working, around the clock.

The data. 

The agent touched systems holding CUI and, according to WIRED's sources, sent credentials and a network map outside the environment. Whatever the classification label says, that's data loss. As the source put it, "Information from a CUI environment is not supposed to leave that infrastructure."

Each of these risks is manageable on its own. Insiders get access reviews. Data gets classified. But when an insider delegates their access to an agent, they create a path that most controls were never designed to see.

Context Is What Makes Data Sensitive

It's tempting to shrug this off because the data was unclassified. The experts WIRED spoke with explain why that's a mistake.

Former NSA hacker Jake Williams notes that unclassified pieces, combined, "can become classified itself." The location of hazardous materials is one thing. Add details on how the facility is protected, and you have something else entirely.

And the most dangerous thing that left wasn't a document. It was credentials and a network map. Matt Triner of Hunter Strategy says a map helps an attacker "understand where the crown jewels are." Credentials let them walk in. Because code is often built on NIPRNet before it moves to air-gapped networks, a backdoor planted upstream could travel somewhere far more sensitive, and standard code scanning likely wouldn't catch it.

This is why we've argued that internal doesn't mean secure. Most DLP programs are tuned to catch a spreadsheet of customer records leaving the building. Far fewer watch for API tokens, admin credentials, architecture docs and infrastructure maps. In SaaS, that material lives in Google Drive folders, Slack channels, GitHub repos and Confluence pages, and an AI agent with an insider's access can read all of it.

An action alone is simply an action. What makes it risky is the context: who (or what) is acting, whose identity it's using, what data it touches and where that data is going.

Four Governance Failures, One Incident

Underneath the insider-and-agent story are four governance failures we see in SaaS environments every day. None of them are new, AI just makes them faster and harder to see.

1. Shadow AI Met Production Data

The agent ran on a personal account. That means no enterprise agreement, no admin console, no centralized logging and no say over how data is handled or retained. Claude is reportedly barred for DoD staff, and Anthropic's own usage policy prohibits bypassing authentication or security controls. The tool was used anyway, by an insider who had every right to be in the system, just not to bring an unvetted agent along.

That's the defining trait of shadow AI: it isn't stopped by policy alone. In SaaS, it often takes one click. An employee connects an AI assistant to Google Workspace, Microsoft 365 or Slack through OAuth, approves broad scopes and gets back to work. Security never hears about it.

2. The Agent Borrowed a Human Identity

This is the part that should worry security teams most. The agent didn't have its own identity. It logged in as the contractor and committed code under the contractor's name. To every log and control in the environment, the agent's actions looked like a trusted employee doing their job.

We've been saying for a while that AI agents are a new class of identity. In DoControl's own analysis of SaaS activity logs, non-human identities (NHIs) accounted for roughly 70% of SharePoint actions and 40% of Google Drive actions. When an agent hides behind a human's credentials, you can't tell those actions apart, and you can't govern what you can't attribute.

3. Access Far Beyond the Task

Writing and deploying code didn't require admin rights or a path around login controls. But sharing a human's credentials is all-or-nothing. The agent inherited everything the contractor could reach, and then some.

This is the oldest problem in access governance: permissions granted for convenience, never scoped to the job and never revoked. With agents, the blast radius grows, because an agent can act on all of that access at machine speed, around the clock.

4. Nobody Saw the Data Leave

The agent reportedly operated for about two months. During that time, credentials and a network map appear to have gone to an outside website. Nothing flagged it in time.

You can't protect what you can't see. But visibility alone wouldn't have been enough here either. Even after the connection was discovered, sources say the response was slow. Visibility without control is useless. The gap between "we found it" and "we fixed it" is where the damage happens.

The Real Root Cause: Well-Meaning Insiders Under Pressure

It would be easy to blame one careless contractor. We don't think that's the lesson.

Triner put it well: contractors face pressure to deliver faster and cheaper, and that pressure works against security. "That existed before AI, but now it's the same problem on speed." He also points out that organizations are often punished for reporting incidents, which discourages exactly the transparency security depends on.

This contractor almost certainly wasn't trying to cause harm. They were trying to get work done. That's the uncomfortable truth about insider risk: most of it isn't malicious. AI agents don't need bad intent to produce bad outcomes, and neither do the insiders who deploy them.

That's why we don't believe the answer is banning AI. People will route around bans, especially when AI makes them dramatically more productive. The answer is data governance that's automated, context-aware, and fast enough to keep up. 

The goal is not to block collaboration or slow down the business. It's to let employees and contractors use AI freely while keeping control over the data it touches and the identities it acts through.

What Security Teams Should Do Now

You don't need a Pentagon contract to have this exact exposure. Here's where we'd start.

  1. Inventory every AI tool and agent touching your data. That includes OAuth apps, browser extensions, API tokens and service accounts, whether connected through corporate or personal accounts. You can't govern an agent you don't know exists.
  2. Give agents their own identities. No agent should operate on a human's credentials. Separate identities make every action attributable and every permission revocable.
  3. Govern contractors like the insiders they are. Apply the same access reviews, monitoring and offboarding to contractors and vendors as to employees, including the agents and apps they connect.
  4. Scope access to the task, not the person. Grant agents the least access they need, use short-lived tokens and review high-risk OAuth scopes such as full Drive, mailbox or admin access.
  5. Block personal AI accounts from corporate data. Make the sanctioned path the easy path, and close the one-click OAuth door for unapproved tools.
  6. Expand what DLP treats as sensitive. Credentials, secrets, architecture diagrams and network maps belong in the same tier as customer and regulated data.
  7. Watch behavior, not just access. Flag unusual volumes, off-hours activity, new external destinations and agents acting from unexpected locations.
  8. Keep a human in the loop for high-stakes actions. Changing permissions, deploying to production and deleting data should require approval.
  9. Put AI agents in your incident response playbook. Define what counts as an agent-related incident, who owns it and how fast it gets reported. Make reporting easy, not punished.
  10. Automate remediation. Revoking a risky integration or suspending a session shouldn't wait on a ticket queue. Two months is a long time; two minutes should be the goal.

Where DoControl Fits

The incident happened on a government network, not in a SaaS app. But the same pattern of agent, borrowed identity, excessive access and data leaving unseen is playing out in SaaS right now. That's the problem DoControl is built to solve.

  • Discover shadow AI and connected apps. DoControl builds a full inventory of OAuth apps, integrations and the identities behind them across your SaaS environment, including the AI tools nobody approved.
  • Govern insiders and non-human identities together. Employees, contractors, vendors, AI agents and service accounts are all part of the same access picture. DoControl ties each agent and app back to the insider who connected it, with risk scored on context, not just activity.
  • Map exposure and sensitivity. See which data is exposed, to whom and how sensitive it is, so you know where an agent could do the most damage.
  • Catch insider and agent risk before it becomes data loss. An agent pulling unusual volumes of data, acting from an unexpected location or sending data to a new external destination gets flagged for what it is.
  • Remediate automatically. Prebuilt and custom workflows can revoke high-risk OAuth integrations, remove unauthorized access and suspend sessions during anomalous activity, without waiting on a ticket.

{{cta-1}}

Final Thoughts

The question isn't whether employees will connect AI agents to sensitive systems. They already are, often with the best intentions and the worst possible access. If it can happen on a heavily secured Pentagon network, it can happen in your Google Workspace.

Every insider who connects an AI agent creates a new identity with access to your data. The organizations that govern those agents like identities, with their own credentials, scoped access, continuous monitoring and fast remediation, will be the ones that get AI's speed without inheriting its risk.

The worst time to discover an AI agent in your environment is after it's already sent your data somewhere else. 

Start with a Free Risk Assessment to see which AI tools and identities can reach your SaaS data today.

Sources

Melissa leads DoControl’s marketing and content strategies, creating educational and engaging narratives that position the brand at the center of the SaaS security market. She translates complex industry trends and security challenges into clear, practitioner-focused insights that highlight DoControl’s unique value.

Her work spans content, campaigns, and brand, connecting strategy and execution across channels to strengthen positioning, inform the market, and shape how organizations think about and approach SaaS security today.

See who (and what) is accessing your data.

Book a demo and see our market-leading DLP platform in action.

Get updates to your inbox

Our latest tips, insights, and news
Tablet top edge with front camera and purple slider control with four dots.