This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
The acting director of CISA uploaded sensitive government contracting documents marked 'for official use only' into a public instance of ChatGPT. This action triggered automated security alerts within the Department of Homeland Security, leading to an internal review regarding the potential exposure of sensitive information. The incident highlights the risks of using public AI tools for handling sensitive government data.
Madhu Gottumukkala, acting director of the Cybersecurity and Infrastructure Security Agency (CISA), reportedly uploaded government contracting documents marked "for official use only" into a public version of ChatGPT. The uploads reportedly triggered automated cybersecurity alerts and prompted a Department of Homeland Security review to assess potential exposure of sensitive information.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI systems that memorize and leak sensitive personal data or infer private information about individuals without their consent. Unexpected or unauthorized sharing of data and information can compromise user expectation of privacy, assist identity theft, or cause loss of confidential intellectual property.
Also classified as
The acting director inappropriately relied on a public chatbot to process sensitive government documents, bypassing standard security protocols.
The incident involved uploading sensitive government contracting documents marked 'for official use only' into a public AI system, creating a risk of unauthorized exposure of non-public sensitive information.
Human
Due to a decision or action made by humans
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The risk was caused by a human decision to upload sensitive documents to a public AI tool, and the potential exposure of data was an unintentional outcome of trying to use the tool for work.
The acting director of CISA uploaded sensitive, non-public government contracting documents to a public instance of ChatGPT, triggering automated DHS security alerts. While no classified information was exposed, the incident underscores the ongoing, established risk of insider data leakage through public AI tools and the challenges of enforcing AI security policies at high levels of government.
The report describes ChatGPT as a 'popular AI tool' and 'app' used by the acting director to process documents, which fits the definition of an Assistant that augments human capabilities but requires human input and oversight.
Limited Risk: The report describes the use of ChatGPT, which is a chatbot. Under the EU AI Act, chatbots are classified as Limited Risk (Level 3) and are subject to transparency obligations.
ChatGPT is primarily designed and used as a writing assistant and question answering tool to help users process, summarize, or draft text and documents.
The report clearly outlines the specific actions taken by the CISA acting director, the nature of the documents uploaded, and the security alerts triggered. While the final outcome of the internal review is unknown, the core facts of the AI safety incident are well-documented by multiple sources.
The incident could be viewed purely as a human policy violation or insider threat rather than an AI safety incident, though the specific data-sharing mechanics of public LLMs are central to the risk.
Technology
Public ChatGPT Version
Public tool shares uploaded data with OpenAI, risking exposure.
Lack of Local Network Controls
Public version lacked federal network containment features of DHSChat.
Data Inputs
Sensitive Document Input
Documents marked 'for official use only' were uploaded into the tool.
Human Factors
User Policy Non-Compliance
Official ignored training on handling sensitive government documents.
Insistence on Tool Access
Official pressured CISA to grant an exception for public ChatGPT.
Process and Methods
Inadequate Exception Process
Exception was granted without enforcing secure, isolated AI alternatives.
Delayed Response to Warnings
Security sensors flagged uploads but did not automatically block them.
Management
Leadership Bypass of Policy
Senior official bypassed default block on ChatGPT for personal use.
Inadequate Oversight of Exception
Management approved the exception without proper risk assessment.
Regulatory Environment
AI Adoption Pressure
Executive order to remove AI barriers pressured rapid tool adoption.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
