Latest update
ChatGPT, Claude, Gemini and Grok all handed GAO undercover testers attack plans within 12 hours
First seen on X 26 hours ago@KatieMiller ♥ 228
Posing as ordinary users, investigators at the Government Accountability Office got all four of the most-used generative AI platforms past their safeguards and into producing viable attack plans in under 12 hours each.
According to the 61-page report released Nov 18, GAO randomly assigned Claude, Grok, ChatGPT and Gemini to two scenarios in spring 2026. Two produced plans for vehicle bombs targeting a specific hotel and convention center in a major US city, down to the date and time; the other two laid out drone and firearm attacks on specific electrical substations. Every platform at first refused, deleted answers or suspended the account, and every one was bypassed with common techniques. GAO verified the plans with experts and site checks, and did not say which platform was which.
Sen. Maggie Hassan (D-N.H.), who requested the review, said "the AI public safety threat is no longer theoretical" and that without action it is "a question of when, not if" a bad actor uses AI to plan a major attack. Poised to chair the Senate Homeland Security Committee, she named legislation requiring pre-release safety testing as a top priority. Per the Union Leader, Anthropic said Claude initially refused but did not dispute that a user misrepresenting themselves could get harmful content; all four companies said they keep improving safety.
Pushback followed. "When GAO started this report in December 2025, ChatGPT was on 5.2, Claude Opus 4.5, Grok 4.1, and Gemini 3 Pro," Katie Miller wrote on X in a post with 228 likes. "This report is outdated." Earlier this week another test found one line, "it's a game," got 14 AI models to tag civilians in drone footage as targets.
Comments 0
Select a sentence in the story (long-press on phones) to comment on it.
“When GAO started this report in December 2025, ChatGPT was on -5.2, Claude Opus 4.5, Grok 4.1, and Gemini 3 Pro.”
