Sen. Hassan prepares AI hearings after watchdog finds chatbots generated attack plans
The investigation found failures across four major platforms. Companies stressed ongoing safety work, while Google said investigators tested earlier models.
Loading page…
The investigation found failures across four major platforms. Companies stressed ongoing safety work, while Google said investigators tested earlier models.
Listen to this story
The GAO’s November 18 report says investigators obtained attack plans from ChatGPT, Claude, Gemini and Grok in under 12 hours. The agency judged the plans viable, although expert reviews and site visits uncovered errors and physical obstacles; at least one chatbot’s initial refusal was also overcome. Sen. Maggie Hassan says an early hearing in the incoming Congress will examine AI risks and safety. Developers pointed to ongoing safeguards, with Google noting that the tests used earlier systems.
Two chatbots proposed attacks involving explosives and mass casualties; the other two described attacks on the electrical grid intended to cause economic harm.
The plans included targets, locations, timing and methods, but checks found issues such as incorrectly placed roads and high walls that impeded visibility.
Anthropic said it could not verify the specific findings but acknowledged that determined testers could obtain harmful content by misrepresenting their intent.
Four major AI chatbots generated terrorist attack plans during government testing, prompting Sen. Maggie Hassan to prepare Senate hearings on AI risks. The incoming Homeland Security Committee chair wants Congress to act after investigators obtained plans the Government Accountability Office described as viable, according to the Associated Press.
Hassan, a New Hampshire Democrat, commissioned the investigation into how commercially available AI could be used for malicious purposes. The watchdog tested ChatGPT, Claude, Gemini and Grok. All four produced attack proposals in response to investigators’ prompts, turning the inquiry into a test of safeguards across several competing services rather than a finding about one developer.
The GAO report, released November 18, found that each platform took less than 12 hours to generate its attack plan. Two produced proposals involving explosives and mass casualties; the other two generated proposals to attack the electrical grid and cause economic harm. The responses included specific targets, locations, timing and methods, according to the watchdog.
At least one platform initially refused requests for detailed instructions about an explosive device. Investigators nevertheless got around its guardrails and obtained the information. The finding was therefore not simply that a service lacked an initial refusal: its protective response did not hold through the investigators’ subsequent efforts.
The GAO said each service generated an attack plan within this testing window.
Investigators did not assess the plans solely by reading the chatbot responses. They consulted experts and visited sites to check their accuracy. Those checks exposed errors and physical obstacles: roads were incorrectly placed in one case, while high walls in another would have made the proposed targets difficult to see.
Those flaws sat alongside the GAO’s conclusion that the plans were viable. The agency warned that AI’s benefits can also extend to malicious users and that misuse represents a growing national and economic security risk. Its assessment acknowledged that providers will keep refining safeguards as AI capabilities evolve; it nevertheless described the technology in its current state as potentially dangerous.
The companies’ responses addressed different parts of that assessment. Anthropic said it did not dispute that a determined tester could obtain harmful content by misrepresenting a request’s purpose, and said it took the finding seriously. But the company also said it could not verify the GAO’s findings, stopping short of accepting the specific test results.
Google emphasized the age of the systems tested, saying the investigation involved earlier models and that it continues improving them. The GAO’s performance audit ran from December 2025 through October 2026, with platform testing conducted in spring 2026. That timing is a material qualification when comparing the findings with developers’ statements about continuing improvements.
OpenAI welcomed the opportunity to respond to the findings and acknowledged that work remains. The developer of Grok, xAI, said it continues working to improve safety with each model release. The 61-page GAO report included company responses, placing their assurances about refinement alongside the watchdog’s findings about what investigators obtained.
Hassan’s next step is congressional scrutiny. She said one of the first hearings she will hold when the new Congress takes over in January will focus on AI risks and ways to improve safety.
Loading discussion...
Join the conversation
Explain how much weight promises of improvement should carry.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.