Tools / News / The Head of U.S. Cybersecurity Uploaded Sensitive Government Files to Public ChatGPT. DHS Banned Commercial AI Across the Agency.
Press

The Head of U.S. Cybersecurity Uploaded Sensitive Government Files to Public ChatGPT. DHS Banned Commercial AI Across the Agency.

· VaultTools

Acting CISA director Madhu Gottumukkala uploaded at least four 'For Official Use Only' documents to public ChatGPT in 2025, triggering DHS cybersecurity alerts and an internal investigation. He was removed in February 2026. DHS has since banned ChatGPT, Claude, and all commercial AI tools for staff.

VaultTools · March 20, 2026

Abstract visualization of artificial intelligence, representing the risks of uploading documents to AI platforms. Photo by Steve Johnson on Unsplash

Table of Contents


What Happened

Between mid-July and early August 2025, Madhu Gottumukkala, acting director of the Cybersecurity and Infrastructure Security Agency (CISA), uploaded at least four documents marked “For Official Use Only” to OpenAI’s public ChatGPT platform. The uploads were flagged automatically by DHS cybersecurity monitoring systems and triggered an internal investigation involving senior Department of Homeland Security leadership.

Politico broke the story on January 28, 2026. Security researchers and members of Congress reacted with alarm: the acting head of the agency responsible for protecting U.S. government infrastructure had done exactly what CISA had warned employees not to do.

Why CISA Had Already Restricted ChatGPT

CISA had blocked access to ChatGPT for the majority of its employees before Gottumukkala’s uploads occurred. The stated rationale was that files and text submitted to the public version of ChatGPT can be retained by OpenAI on its servers and potentially incorporated into model training data, making sensitive content accessible to hundreds of millions of users worldwide.

Gottumukkala had personally requested and received a special exemption from this restriction. The documents he uploaded were not classified but carried a “For Official Use Only” designation, a category intended for information that, while unclassified, carries a risk of harm if disclosed.

The Investigation and Its Outcome

DHS launched an internal investigation after the automated alerts. Senator Chuck Grassley sent a formal letter to CISA on February 5, 2026, requesting a full account of the uploads and any remediation steps taken. Gottumukkala was removed from his acting director role on February 27, 2026.

The DHS-Wide Commercial AI Ban

Following the incident, the Department of Homeland Security directed staff to stop using commercial AI tools including ChatGPT and Claude for work purposes. Employees were redirected to DHSChat, an internal AI platform built on models that do not retain or externally expose submitted content. The ban applies across DHS components, not only CISA.

The shift reflects a policy reality that many enterprises and governments have been moving toward: general-purpose commercial AI tools with cloud-side processing are incompatible with handling sensitive documents, regardless of how the tool is positioned.

What the Incident Illustrates About File Uploads

The CISA episode is notable because it involves not a careless junior employee but the acting head of the agency whose mission is federal cybersecurity. It demonstrates that the risk of cloud-side file processing is not theoretical and does not require a breach: the act of uploading a file to a cloud service is itself the exposure event.

When a file is submitted to any cloud-based tool, whether an AI assistant, a PDF converter, an image compressor, or a document translator, it travels to a remote server. What that server retains, logs, or does with the content is governed entirely by the service’s terms and infrastructure. The user has no visibility and no control once the file leaves the device.

Browser-based tools that process files using WebAssembly never receive the file on any server. The document is read from disk, processed inside the browser tab, and written back to disk. There is no upload, no retention policy, and no server to monitor or audit. It is the architecture the DHS policy is now trying to replicate through internal tooling.


Sources