AI Is Going Just Great
← Timeline
·4d agoScaryMajor

GAO: All Four Major Chatbots Produced Workable Terror Attack Plans When Asked

Published · updated · curated by AI Is Going Just Great

Source: fortune.com ↗ · archived copy ↗

"We got the nefarious information that we were after. So, in this case, those safeguards didn't work." — Kevin Walsh, GAO director of IT and cybersecurity

Government Accountability Office investigators posing as would-be attackers got ChatGPT, Claude, Gemini, and Grok to produce viable plans for a bombing and a power-grid attack, each in under 12 hours. Every chatbot's safeguards fired; testers bypassed them "by using common safeguard-bypassing techniques," GAO wrote. Two platforms gave step-by-step IED plans targeting a hotel parking garage and a convention center. The other two named specific drone and firearm models for attacking electrical substations and explained how to buy them.

The caveats are real but not reassuring. GAO tested each platform on a single scenario, won't say which chatbot produced which plan, and found some details were wrong — one map placed roads in the wrong locations. OpenAI told ABC the test used "an older version of ChatGPT"; Google said it ran on "earlier models." Anthropic did not dispute "a determined tester can get harmful content" from its system. GAO made no recommendations.