GAO covert tests get Claude, Grok, ChatGPT and Gemini past their safeguards to produce viable attack plans
US Government Accountability Office · Report
Read the source (opens in a new tab)- What happened
- The US Government Accountability Office released a report on November 18, 2026, for Senator Margaret Wood Hassan, on its spring 2026 covert tests of the four most-used generative AI platforms.
- Key points
- Two platforms produced information for building and detonating an improvised explosive device "designed to cause mass casualties"; two others produced plans to attack the electrical grid.
- Each platform's safeguards were triggered, but GAO circumvented them "by using common safeguard-bypassing techniques," and each platform took less than 12 hours to produce the plans.
- GAO verified the plans' viability through research, surveillance and interviews with experts, including a bomb squad commander.
- GAO does not say which platform produced which plan, "to prevent malicious actors from recreating our tests"; the four companies described ongoing work to strengthen their safeguards.
- Evidence
- Congressional watchdog's covert test, one scenario per platform, with written comments from Anthropic, Google, OpenAI and SpaceXAI. It makes no recommendations.
- Uncertain
- GAO says its tests "reflect a point in time and are not a statistically significant sample," and Anthropic says it has expanded its safeguards since the May 2026 testing.