U.S. Sen. Maggie Hassan plans to make artificial intelligence safety one of the first subjects addressed by the Senate Homeland Security and Governmental Affairs Committee when she takes over as chair in the new Congress.
Her announcement followed the public release Wednesday of a 61-page Government Accountability Office audit that examined whether widely used commercial AI systems could be misused for attacks in the United States. Investigators tested ChatGPT, Claude, Gemini and Grok in covert exercises during 2026 and obtained information that GAO said could help construct plans for mass casualties or economic harm.
“Congress must quickly take up and pass legislation to address AI safety threats. Today’s new report makes it clear that the AI public safety threat is no longer theoretical. 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.”
— Sen. Maggie Hassan
Safeguards were bypassed
GAO said each platform initially activated protections such as refusing requests, removing generated material or suspending accounts. Investigators then bypassed those safeguards and obtained usable information from all four systems in less than 12 hours of testing. The agency said the time required could vary depending on the subject and the user’s experience.
The tests assigned two platforms to a scenario involving an explosive device intended to cause mass casualties. The other two were tested on a scenario involving an attack on electrical infrastructure to produce economic damage. GAO said investigators used responses from the platforms along with outside research and expert consultations to evaluate whether the resulting plans were workable. The agency did not publish the prompts or the methods used to defeat the safeguards.
- GAO tested four widely used, U.S.-based platforms selected using survey and web-traffic data.
- The exercises were conducted in spring 2026 as part of an audit that ran from December 2025 through October 2026.
- GAO said the testing involved individual scenarios and was not a statistically significant sample of all possible uses.
GAO said the results showed that generative AI systems could provide information useful to malicious actors even after their protections were triggered. The agency also said the success of a real-world attack would depend on circumstances beyond the information supplied by a platform.
Hassan requested the investigation while serving as ranking member of the Joint Economic Committee. She said the committee’s early work in the next Congress will include a hearing on AI risks to homeland security and critical infrastructure, as well as consideration of legislation addressing the safety concerns identified in the audit.
The companies connected to the platforms told GAO that they were continuing to improve their security measures. GAO wrote that the technology’s benefits are substantial but that its capabilities can also aid malicious actors, creating risks to national and economic security.