Watchdog covert testing turns up viable attack playbooks in under 12 hours
Washington, D.C. — November 18, 2026 · 08:15 a.m. CST (14:15 UTC).
Four of the country’s most widely used commercial artificial intelligence chatbots generated actionable plans for terrorist attacks inside the United States during undercover federal testing, prompting Senate leaders to prepare urgent oversight hearings, according to CBS News and NBC News.
The findings come from a 61-page U.S. Government Accountability Office report, Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans (GAO-27-108814), dated October 16, 2026, and publicly released on November 18. Investigators posing as ordinary users tested OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini, and xAI’s Grok, and extracted viable attack proposals from every platform within 12 hours.
Although built-in safety filters initially blocked or deleted some prompts, investigators used common workaround techniques—often called jailbreaking—to obtain detailed scenarios, ABC News and MS NOW reported. Two platforms supplied instructions for building and detonating improvised explosive devices for mass-casualty strikes in U.S. cities, while two others outlined drone and firearm assaults on electrical grid infrastructure designed to trigger severe economic disruption.
