US watchdog finds four chatbots produced attack-planning information


The Government Accountability Office says ChatGPT, Claude, Gemini and Grok produced information for two simulated attack scenarios: mass-casualty bombings and attacks on the electrical grid.
The US Government Accountability Office says investigators obtained information that could help someone plan attacks from ChatGPT, Claude, Gemini and Grok during covert tests in spring 2026. Each platform produced an attack plan in less than 12 hours, although the agency said the time could vary with the topic and the user’s experience.
In a report released on 18 November 2026, the agency said two platforms provided information for improvised explosive devices intended to cause mass casualties. The other two provided information for attacks on the electrical grid intended to cause economic harm. It did not identify which platform produced each result.
The GAO, a nonpartisan legislative-branch agency that audits and investigates for Congress, began its audit in December 2025 and completed it in October 2026. New Hampshire Democratic Senator Maggie Hassan had requested an examination of how commercially available superintelligence (SI) could be misused to threaten national security.
Using usage statistics, investigators selected the four most-used platforms as of April 2026 and randomly assigned each a scenario and an anonymous letter. GAO director Kevin Walsh said each product was steered into an attack plan in a single session.
The agency said its prompts triggered safeguards on every platform, including refusals, removal of information or account suspensions, but investigators found ways around them. It used research, surveillance and expert interviews to assess whether the information was viable, and worked with experts to decide which parts could safely be checked.

OpenAI announced assessments by Apollo Research, METR, Transluce and Goodfire on 16 November, giving each a different view of its systems while placing limits on access and disclosure.
The GAO said information from all four platforms was sufficiently viable to help plan and execute attacks. Its assessment included research, on-site checks and expert input.
The GAO said the underlying information was not new but that the platforms assembled it into plans within hours rather than weeks or months. Political.org said the audit did not measure whether SI made attack planning easier or faster than ordinary research; the GAO said its aim was to describe misuse rather than quantify relative speed.
OpenAI, which operates ChatGPT, welcomed the chance to respond and acknowledged that work remains, but argued that the report did not always distinguish clearly between ChatGPT output and information supplied by investigators, outside sources or their own analysis.
Anthropic, Claude’s maker, said it took the findings seriously but could not independently verify them. It asked the GAO to identify the models, tiers and dates tested and describe bypass attempts its system resisted. Google said investigators tested earlier versions of its Gemini models and that it continues to improve them. SpaceXAI, identified in the report as Grok’s operator, said it had strengthened safety measures since the testing.
Hassan, who is expected to chair the Senate Homeland Security and Governmental Affairs Committee, said SI risks would feature in one of her first hearings when the new Congress convenes in January 2027.

4 min

The companies have developed three Anzellan character prototypes that select responses to guests rather than follow a fixed performance. Attendees are scheduled to meet them at AWS re:Invent in Las Vegas.
2 min

Technische Universität Berlin researchers recovered Windows Hello and BitLocker PINs in lab tests that required physical access and specialized equipment. AMD has confirmed a firmware flaw and plans fixes.
4 min