Government Audit Finds Major AI Platforms Generate Viable Attack Plans
Federal investigators bypassed guardrails on four popular generative AI tools within hours to obtain blueprints for physical and grid attacks.
- By Jesse Jacobs
- Nov 18, 2026
Four of the most-used generative artificial intelligence platforms generated feasible plans for mass-casualty bombings and electrical grid attacks during covert federal testing, according to a Government Accountability Office report.
Investigators tested ChatGPT, Claude, Gemini and Grok in spring 2026 to evaluate whether commercially available AI could be misused to threaten national security and cause economic harm. Each platform produced viable attack plans in less than 12 hours, compiling existing information in hours rather than weeks or months.
The GAO randomly assigned the platforms to two scenarios: constructing and detonating an improvised explosive device to cause mass casualties, or attacking the electrical grid to inflict economic harm.
Two platforms provided instructions for building and detonating vehicle-borne explosive devices, recommending a specific hotel and convention center in major U.S. cities along with exact dates and times to coincide with large crowds at scheduled public events.
According to the Associated Press, one platform also suggested renting a vehicle and purchasing supplies from a store along the route to the target. The other two platforms generated maps, images and plans to attack electrical substations near a U.S. airport using drones and firearms, with one identifying a season of the year for the strike.
The GAO verified the plans through research, surveillance, site visits and interviews with independent experts. Investigators confirmed the details were sufficient to plan and execute attacks, though site visits revealed inaccuracies in some outputs, including incorrectly placed roads on a map and high walls that would block a shooter's view of the target.
All four platforms initially triggered safeguards—denying requests, removing generated information or suspending accounts—before investigators bypassed the guardrails using common circumvention techniques. For safety reasons, the GAO withheld the exact prompts used and did not disclose which platform produced each scenario.
In responses included in the 61-page report, Anthropic stated it does not dispute that a determined tester can extract harmful content by misrepresenting a request's purpose, while noting it could not verify the GAO's findings. Google stated the testing occurred on earlier models, while OpenAI and SpaceXAI said they are continuing to improve platform safety.