Timeline

GAO finds four chatbots gave workable attack plans in covert tests

Prompts in spring 2026 produced plans for a vehicle bomb or substation attacks within 12 hours per platform; GAO checked viability with experts and did not say which chatbot produced which plan.

  • Safety & alignment
  • Government & policy
  • Notable

The US Government Accountability Office, the investigative arm of Congress, reported that prompts to four widely used chatbots in spring 2026 produced information that could help plan an attack. GAO tested Claude, Grok, ChatGPT and Gemini, which its usage statistics identified as the four most-used generative AI platforms as of April 2026. Sen. Maggie Hassan, who requested the work, is due to chair the Senate Homeland Security Committee from January.

GAO selected two scenarios with expert input: an improvised explosive device causing mass casualties, and an attack on the electrical grid causing economic harm. Two platforms produced information for building and detonating a vehicle-borne device; two produced plans for attacking electrical substations, one by drone and one with firearms. GAO said each response included specifics such as targets, locations, timing and methods, that each platform took less than 12 hours, and that it verified viability through research, surveillance and expert interviews. Every platform’s safeguards triggered at some point, through denial, removal of content or account suspension, and GAO bypassed them on all four using what it called common techniques.

Limits apply. GAO randomly assigned the platforms to scenarios and labelled them A to D, so the report does not say which chatbot produced which plan. The Associated Press reported that GAO’s checks also found errors, such as wrongly placed roads. The tests used spring models; Google said the testing was on earlier models, and Anthropic’s letter in the report said it does not dispute that a determined tester can extract harmful content by misrepresenting a request, adding that it takes the finding seriously. OpenAI said it welcomed the chance to respond, and SpaceXAI said it “continues to work diligently” on safety.

Hassan said one of her first hearings in the new Congress would address AI risks and safety improvements, adding, “If Congress does not act, it is a question of when — not if — a bad actor will use AI to plan and carry out a major attack.”

A separate GAO technology assessment, AI and Cybersecurity, lays out four policy goals and 16 options for policymakers. It notes agentic AI being used to carry out several stages of an attack with minimal human input, and lists options including requiring human authorisation for consequential AI-initiated cybersecurity actions.

In the commentary

What people were saying around this time — external links, from the record's commentary rail.