WASHINGTON: The US Government Accountability Office (GAO) found that four widely used artificial intelligence (AI) chatbots supplied feasible attack-planning information after investigators bypassed their safeguards, according to a report released Wednesday.
The investigation tested ChatGPT, Gemini, Claude and Grok during spring 2026. All initially triggered protections, including refusals, information removal or account suspensions, but investigators eventually obtained harmful responses.
Two platforms supplied information concerning potential mass-casualty attacks. The others generated material about attacks on electrical infrastructure intended to cause economic harm.
GAO checked the information through research, site assessments and consultations with experts. Each platform produced its responses in less than 12 hours, although the agency cautioned that completion times could vary.
Investigators withheld the prompts, bypass techniques and links between particular models and scenarios to prevent others from recreating the tests. The findings also carry methodological limits.
GAO acknowledged that the tests represented a point in time rather than a statistically significant sample. Because investigators assigned different scenarios to different platforms, the results do not support comparisons between the chatbots. Companies challenged aspects of the assessment while describing continuing safety work.
Anthropic, which operates Claude, questioned the sample size and absence of a control group. It maintained that its safeguards had continued to develop since testing.
OpenAI said ChatGPT had refused numerous requests or offered safer alternatives. It also questioned whether the report consistently distinguished the chatbot’s contributions from information supplied by investigators or external research.
Google described safeguards throughout Gemini’s development process and continuing model improvements. Grok operator SpaceXAI said it was working to strengthen safety and moderation with successive releases.
Read More: Optimizely Fixes Flaw After Malware Targets Web Visitors
GAO nevertheless maintained that all four platforms ultimately supplied information that malicious users could exploit.
Democratic Senator Maggie Hassan of New Hampshire requested the investigation. She argued that the findings demonstrated the need for legislation addressing AI risks to public safety and critical infrastructure.
“The AI public safety threat is no longer theoretical,” she said.
Hassan, the incoming chair of the Senate Homeland Security and Governmental Affairs Committee, is considering stronger safety-testing requirements before public model releases. She plans to make AI security an early hearing priority when the new Congress begins in January.