Government investigators obtained information they assessed could facilitate attacks from ChatGPT, Claude, Gemini and Grok, despite safeguards intended to stop harmful requests, according to a Government Accountability Office report released Wednesday.

GAO examined scenarios involving national-security and economic harm. The report keeps the individual platforms anonymous by scenario.

The tests took place this spring. Investigators obtained attack-planning information from each platform in less than 12 hours, although GAO said the time required could vary with the subject and the user’s experience.

The platforms refused requests, removed information or suspended accounts during testing. Investigators nevertheless got around those protections. GAO consulted experts to assess whether the information could facilitate attacks.

Sen. Maggie Hassan, the incoming chair of the Senate Homeland Security and Governmental Affairs Committee, said AI safety would be an early hearing priority when the new Congress convenes. She is also considering legislation requiring stronger testing before models are released. Her announcement did not specify a hearing date.

The companies emphasized continuing safety improvements. Google noted that earlier models were tested. Anthropic said it took the findings seriously but could not verify them. OpenAI acknowledged work remained, while SpaceXAI said it was improving safeguards with each release, AP reported.