ai-cybersecuritytrade-pressNewsThe Broadside1 min read

GAO finds ChatGPT, Claude, Gemini, Grok generated viable attack plans

All four platforms' safeguards fell to routine bypass techniques in under 12 hours each, investigators found.


TL;DR

Covert GAO testing in spring 2026 found ChatGPT, Claude, Gemini, and Grok each produced viable, expert-verified plans for a mass-casualty IED and an electrical-grid attack. Safety guardrails at first refused requests or suspended accounts, then folded under common bypass techniques, all in under 12 hours per platform. The report withholds which platform produced which scenario and the specific techniques. GAO issued no recommendations; Sen. Maggie Hassan, who requested the review, says AI safety legislation is now an early priority.

The guardrails on the four most widely used AI platforms failed under routine pressure, not exotic attack. That's the blunt finding of a Government Accountability Office report published Wednesday: undercover investigators coaxed ChatGPT, Claude, Gemini, and Grok into producing viable, expert-verified plans for two kinds of attacks, a mass-casualty IED and a strike on the electrical grid, in under 12 hours apiece.

No sophisticated jailbreak was required. GAO's Kevin Walsh, a director of IT and cybersecurity, was direct: "We got the nefarious information that we were after. So, in this case, those safeguards didn't work. And I think that's the story here." Investigators used common safeguard-bypassing techniques; the platforms at first refused requests, removed content, or suspended accounts, then folded anyway.

The plans included specific targets, locations, timing, and methods, and GAO worked with experts to confirm the information could actually be used. Walsh named the companies: "Google, Anthropic, OpenAI and SpaceXAI... the product that they're selling is potentially dangerous."

One asymmetry sits in the report itself. It does not say which platform produced which scenario, and it deliberately withholds the bypass techniques, to stop malicious actors from recreating the tests. The result is a blanket finding of failure without a public accounting of which safeguards gave way.

GAO issued no recommendations, since it was testing third-party products rather than federal agencies, and conceded the sample is not statistically significant. The legislative push now falls to Sen. Maggie Hassan, D-N.H., who requested the review. She says AI safety legislation and a hearing on homeland-security risks will be an early priority when she chairs the Senate Homeland Security and Governmental Affairs Committee next Congress, and she is weighing tighter safety testing before public release of models.


Published ·Deep Fathom

GAO finds ChatGPT, Claude, Gemini, Grok generated viable attack plans — The Broadside