ChatGPT, Claude, Gemini And Grok Helped Plan Attacks In Under 12 Hours. US Tests Expose Safety Gaps.
ChatGPT, Claude, Gemini and Grok provided information that could help plan deadly attacks in the United States in less than 12 hours each, despite safeguards designed to prevent such assistance, according to a federal investigation released Wednesday. The findings have prompted calls for congressional action as lawmakers question whether leading artificial intelligence companies are doing enough to prevent their technology from being misused.
The Government Accountability Office report found that investigators were able to circumvent safety protections on all four platforms and obtain information assessed as feasible for planning attacks involving mass casualties or damage to critical U.S. infrastructure.
The tests were conducted in spring 2026 using versions of the platforms available at the time. Although the companies have since strengthened their safeguards, the findings highlight the challenges of preventing determined users from obtaining potentially dangerous information.
Two platforms generated information that could assist in planning explosive attacks in major American cities, while the other two produced material related to attacks on the U.S. electrical grid that could cause significant economic damage.
The GAO did not disclose which chatbot was assigned to each scenario, citing concerns that identifying the platforms and their responses could make it easier for others to reproduce the results.
All four initially activated safety protections, including refusing requests or restricting certain information. However, investigators eventually obtained material that the agency determined could assist someone attempting to carry out an attack.
The watchdog verified portions of the AI-generated information through independent research, observations at relevant locations and consultations with specialists.
Investigators concluded that the information was largely viable, although some responses contained inaccuracies, including incorrect geographical details and assumptions about conditions at particular locations.
No actual attacks were carried out during the tests.
One of the report’s most significant findings was how quickly investigators obtained the information.
Each platform produced material that could assist attack planning in less than 12 hours. The GAO cautioned that the time required could vary depending on the subject, the user’s experience and other factors.
The agency also acknowledged that much of the underlying information was already publicly available. However, it found that AI systems could organize existing material into plans within hours, potentially accelerating research that might otherwise take weeks or months.
The investigation was requested by Sen. Maggie Hassan, D-N.H., who is preparing to lead the Senate Homeland Security and Governmental Affairs Committee when the new Congress convenes in January.
In a statement released Wednesday, Hassan warned that the findings demonstrated the need for stronger AI safeguards and congressional oversight.
“The AI public safety threat is no longer theoretical,” Hassan said.
She announced that one of the committee’s first hearings in the new Congress would examine the risks AI poses to national security and critical infrastructure.
Hassan also said lawmakers would explore legislation addressing those risks, including more rigorous safety testing before companies release new AI models.
The findings come as technology companies compete to develop increasingly capable AI systems that can process complex requests, conduct research and perform tasks with limited human assistance.
OpenAI, Anthropic, Google and SpaceXAI submitted responses to the GAO, outlining their existing safeguards and efforts to improve protections against misuse.
OpenAI said it takes seriously the possibility that AI could facilitate violence and described its use of safety testing, automated detection, human review and other measures.
The company also questioned aspects of the investigation, arguing that the report did not always clearly distinguish information generated by ChatGPT from material supplied by investigators or obtained through outside research.
The GAO acknowledged that its investigators used their own experience and consulted external sources during the tests. However, it maintained that ChatGPT ultimately provided information that could assist an attack.
Anthropic similarly raised concerns about the study’s methodology, noting that the limited number of tests was insufficient to produce statistically significant conclusions about the overall reliability of AI safety systems.
The company also argued that a comparison with conventional research methods would help establish how much additional capability AI provided to someone seeking dangerous information.
The GAO acknowledged those limitations but said its objective was to demonstrate how commercially available AI could be misused, rather than measure whether the platforms made harmful activities easier or faster than other methods.
Anthropic also emphasized that the investigation reflected its systems as they existed in May and that it had continued expanding safeguards since then.
Google described its ongoing efforts to strengthen Gemini’s protections through safety evaluations, human review and changes throughout the development process.
SpaceXAI said it was continuing to improve Grok’s safety and moderation systems and requested additional testing information to investigate the findings.
The GAO said it provided the companies with access to the accounts used during testing in September to assist their investigations.
The agency also stressed that its tests represented a limited assessment of the platforms at a particular point in time.
Researchers examined four platforms under selected scenarios rather than conducting a comprehensive comparison of their overall safety performance. The findings therefore do not establish how frequently similar attempts would succeed or whether updated models would produce the same responses.
Nevertheless, the watchdog concluded that the results demonstrated a national security concern, particularly as AI tools become more capable and widely available.
The Associated Press reported that Hassan intends to make AI safety an early priority when she assumes leadership of the Senate committee.