GAO covert tests coax viable attack plans from AI platforms
Technology
Wednesday, November 18, 2026
- Updated Nov 18, 2026 12:06
In covert tests by the Government Accountability Office, all four of the most used generative AI platforms produced information verified as sufficient to plan attacks, including a vehicle-borne bomb and strikes on electrical grid substations, after testers bypassed safeguards with common techniques.
Why it matters: A federal auditor has now demonstrated, and verified with experts, that the leading consumer AI platforms can be steered past their safeguards into producing workable attack plans in under 12 hours, a concrete data point in the debate over AI safety controls.
Data as of US Government Accountability Office report GAO-27-108814, read in full at gao.gov on November 18, 2026; the covert tests described were run in spring 2026 against the four most used platforms as of April 2026, which GAO names in its methodology and anonymises in its results table. Developing item.
In covert tests run in spring 2026, the Government Accountability Office prompted four of the most used generative AI platforms, Claude, Grok, ChatGPT and Gemini, and each produced information GAO verified as sufficient to plan and execute attacks, including constructing and detonating a vehicle-borne improvised explosive device designed to cause mass casualties and targeting electrical grid substations with a drone or firearms. Each platform’s safeguards did trigger, through denials, removal of information or account suspensions, but GAO circumvented them using common safeguard-bypassing techniques, and each platform produced its material in less than 12 hours.
GAO verified the viability of the plans through research, surveillance and interviews with experts, and presented the results with the platforms anonymised. Officials from the four companies generally told GAO their companies will continue to improve the safety of their platforms.
Sources
- Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans (GAO-27-108814) - US Government Accountability Office - 2026-11-18