Four widely used AI chatbots produced information that government investigators judged feasible for planning attacks on people and U.S. critical infrastructure, even after the systems initially triggered their safety protections, according to a newly public report from the U.S. Government Accountability Office.

GAO made the report public on November 18, although it carries an October 16 publication date. It covers covert tests carried out in spring 2026 on ChatGPT, Claude, Gemini and Grok. The agency said two platforms produced information that could help plan mass-casualty attacks using improvised explosive devices, while the other two supplied information for attacks on the electrical grid using drones or firearms. GAO did not identify which platform was associated with each scenario in the public report. The GAO report says each platform generated a viable plan in less than 12 hours.

GAO said all four systems initially triggered safeguards, including refusals, removal of content or account suspensions. Investigators were still able to bypass those protections using common techniques and obtain material that experts judged viable. The agency assessed the results through research, expert input and site visits, while noting that some details supplied by the systems were inaccurate.

The findings are based on models tested in spring 2026 and do not necessarily describe current versions. Anthropic said a determined tester could obtain harmful content by misrepresenting a request, but said it could not verify GAO’s findings. Google said the tests involved earlier models, while OpenAI and SpaceXAI said they continue to improve safeguards. The report includes responses from all four companies.

READ
Jordan Blocks Telegram Nationwide Over National Security Concerns

Sen. Maggie Hassan, who requested the review and is expected to lead the Senate Homeland Security Committee in the next Congress, said she plans to hold hearings on AI safety when the new Congress convenes in January, the Associated Press reported. The findings come as regulators and researchers scrutinize AI misuse, including a California lawsuit over AI cybersecurity risks and research into how AI agents can automate real-world software exploits.


Buy ExpressVPN with PayPal or Credit Card
Advertisement