Skip to content

GAO: four AI chatbots generated terror attack plans in a 12-hour test

Category
News
Published
Reading time
2 min

ChatGPT, Claude, Gemini and Grok gave US government testers terror attack and power-grid plans in 12 hours, a GAO report released November 18 found. Hearings on AI safety were scheduled by Sen. Hassan.

Four AI chatbots produced usable terror attack plans within 12 hours during a covert test requested by Sen. Maggie Hassan, D-N.H., the US Government Accountability Office (GAO) said in a report released Wednesday, Nov. 18.

Investigators got two detailed mass-casualty bombing plans, one for a hotel parking garage and one for a convention center, each with precise timing.

Two other chatbots mapped out attacks on the power grid. One produced a four-step diagram and a simulated video for a drone strike on a substation, and the other gave instructions for disabling a substation near a major airport and data centers with a rifle shot.

Testers posed as ordinary users and got around the chatbots' initial refusals using jailbreak techniques, according to the GAO. Some models at first declined to answer or suspended the test accounts before investigators found ways around the safeguards.

Hassan called the result “a concrete playbook for how to carry out a deadly attack” in under 12 hours. She added that the danger such tools pose to public safety “is not theoretical.”

OpenAI called the findings “concerning” but said the test ran on an older ChatGPT version, adding that it continually checks its safeguards with internal and outside experts. Google said it keeps strengthening its product protections. Anthropic acknowledged that a determined user can still extract harmful material from Claude despite its safeguards.

The GAO did not disclose which chatbot produced which response.

Hassan takes over as chair of the Senate Homeland Security Committee in January 2027 and has already scheduled hearings on AI safety.