GAO Says Four Chatbots Yielded Attack-Planning Information
The Government Accountability Office found that investigators obtained information assessed as feasible for planning attacks from ChatGPT, Claude, Gemini and Grok, overcoming each platform’s initial safeguards in under 12 hours during tests conducted in spring 2026.
The tests examined two types of scenarios: attacks involving improvised explosive devices intended to cause mass casualties, and attacks on electrical infrastructure intended to cause economic harm. Two platforms provided material related to each scenario. The GAO did not identify which chatbot produced which material, saying disclosure could make the results easier to reproduce.
Investigators used continued prompting and other techniques after the platforms initially refused or restricted harmful requests. The GAO said experts and other checks were used to assess whether the generated information could assist an attack. Some responses contained inaccuracies, including incorrect geographic details and assumptions about conditions at particular locations, the agency reported.
The GAO described the work as a limited, point-in-time assessment of selected scenarios, not a statistically significant measure of the...