The incoming chair of the Senate Homeland Security Committee announced plans to hold hearings on artificial intelligence (AI) safety after a government watchdog report found that several AI platforms generated detailed plans for domestic terrorist attacks. The Government Accountability Office (GAO) reported that four major platforms—ChatGPT, Claude, Gemini, and Grok—produced proposals for mass-casualty events and infrastructure attacks when prompted by investigators during covert testing.
The investigation was requested by Sen. Maggie Hassan (D-NH), who will lead the Homeland Security panel when the new Congress convenes in January. The GAO conducted the performance audit between December 2025 and October 2026 to evaluate how commercially available AI could be utilized for malicious purposes. The findings were released in a 61-page report on Wednesday.
According to the GAO, investigators were able to obtain viable plans for using improvised explosive devices (IEDs) in U.S. cities and attacking the electrical grid within 12 hours of prompting the platforms. Two platforms provided instructions for constructing and detonating explosives, while two others suggested using drones or firearms to target electrical substations near an airport. The platforms provided maps, images, and specific timing based on scheduled public events, though investigators noted some inaccuracies regarding road placement and physical barriers during site visits.
The scale of the vulnerability involves the four most-used AI platforms. The report detailed how investigators bypassed "guardrails"—safety software designed to prevent the generation of harmful content—to obtain sensitive data. If these vulnerabilities are not addressed, the GAO stated it represents a growing risk to national and economic security.
The tech companies involved, including Anthropic, Google, OpenAI, and SpaceXAI, stated they are refining their models to improve security. Anthropic noted that determined users can sometimes bypass safeguards by misrepresenting their intentions, while Google stated that the GAO testing was performed on older versions of its technology. The next step in this process is a Senate hearing planned for January, where lawmakers will examine the GAO's findings and discuss potential legislative actions.