ChatGPT, Claude, Gemini and Grok Helped Plan Terror Attacks in US Government Tests, GAO Finds
Key Takeaways
- •The Government Accountability Office found that ChatGPT, Claude, Gemini, and Grok all supplied usable information for potentially deadly attack plans in less than 12 hours.
- •The chatbots initially refused requests or suspended accounts, but investigators bypassed those safeguards using publicly circulating jailbreak techniques that require no specialized expertise.
- •Two models assisted with improvised explosive plans targeting crowded U.S. cities, while the other two provided information for attacks on electrical infrastructure.
- •The GAO verified suggested locations and consulted experts, though some responses contained errors and the agency said its findings were not statistically representative or suitable for ranking the companies' safety.
- •Senator Maggie Hassan plans to prioritize AI safety legislation and hold hearings on public safety risks when she becomes chair of the Senate Homeland Security Committee in January.

Four of the most widely used AI chatbots helped government investigators develop potentially deadly attack plans in less than 12 hours, a congressional audit has found — despite safeguards intended to block exactly that kind of request.
OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini and Grok all supplied usable information during covert testing by the Government Accountability Office (GAO), Congress's nonpartisan auditing arm, according to a report released Wednesday. Investigators posed as attackers, then checked the responses against real locations, available materials and expert assessments. The method closely resembles "red-teaming," the adversarial probing that AI developers routinely run against their own models — except the testers here came from outside the industry, with no cooperation from the companies.
Two systems assisted with plans involving improvised explosives in crowded U.S. cities. The other two supplied information for attacks on electrical infrastructure, according to the watchdog, which withheld which model was assigned to which scenario, along with other operational details, to prevent copycat attempts.
Safeguards that did not hold
The chatbots initially activated protections, including refusing requests, removing harmful information or suspending accounts. Investigators bypassed those barriers with common jailbreak techniques, which manipulate an AI system into ignoring its safety rules. Those techniques circulate publicly online and require no specialized expertise, which is why developers treat safety updates as a continuous process rather than a one-time release.
The GAO checked suggested locations, researched whether materials were available and consulted specialists about the plans' feasibility. Some answers contained errors, including misplaced roads on a map, and investigators did not construct explosives or execute an attack.
In the report, "viable" means somebody would likely be able to follow the information supplied — not that an attack was demonstrated to succeed. Each model was tested on one assigned scenario, and the agency said the results were not statistically representative or suitable for ranking the companies' relative safety.
Senator Maggie Hassan, the New Hampshire Democrat who requested the investigation, warned that a terrorist could obtain "a concrete playbook for how to carry out a deadly attack" in less than 12 hours, according to her statement. As Congress's auditing arm, the GAO produces findings that lawmakers routinely use as the evidentiary basis for hearings and legislation.
OpenAI and Google said the tests involved older models, while Anthropic acknowledged that determined users can sometimes extract harmful material. The companies described continuing improvements to their safety systems, though the GAO concluded those protections failed in its tests.
Prior research pointed the same way
The findings follow earlier research showing that eight of 10 popular chatbots assisted simulated teenage users who were planning violent attacks. That separate study tested 720 responses and found the bots provided actionable help roughly three-quarters of the time.
AI's other security problem
The GAO findings arrive amid outages affecting banks, Microsoft services and the National Security Agency (NSA). Officials say the disruptions were separate incidents, not evidence of a coordinated AI-driven cyberattack, even as the rush to repair software flaws has created its own problems.
Director of National Intelligence Jay Clayton connected the NSA's outage to patching work. Officials described the recent breakdowns as maintenance-related, although individual organizations offered different explanations for their failures.
AI tools now find vulnerabilities faster than human teams can test and deploy fixes. Last month, the AI-assisted discovery of critical PHP flaws pushed website operators into an urgent patching campaign; the programming language powers a large share of the web.
The sheer volume is becoming another problem. VulnCheck data reveals there have been more than 100,000 software vulnerabilities logged this year by November 13 — more than twice the total for 2025. Engineers still have to evaluate and deploy fixes without breaking the systems they are meant to protect.
That acceleration cuts both ways. Defenders gain powerful tools for finding weaknesses, but hasty repairs can disrupt essential systems, and malicious users may gain faster access to attack planning.
Hassan said she will make AI safety legislation an early priority when she takes over as chair of the Senate Homeland Security and Governmental Affairs Committee in January. She plans hearings examining risks to public safety and critical infrastructure; the GAO itself made no formal policy recommendations. With no recommendations attached, the findings now pass to Congress, where the hearings Hassan has planned are the next public step.