Story
Nov 21, 2026
AI safeguards failed a terror-plot test, but the findings come with limits
The GAO’s test raises alarms for Sen. Maggie Hassan, who says Congress must act, while AI companies stress that the systems tested were older versions and the watchdog cautions that its findings are limited. The results also sharpen a broader debate over whether faster vulnerability discovery and patching are creating new risks for critical services.
Tests conducted in the spring found that investigators could get ChatGPT, Claude, Gemini and Grok to produce plans for attacks on U.S. targets. Two systems provided information for an improvised explosive device; two others offered ways to target electrical substations. Each plan took less than 12 hours, according to the GAO report, which said the researchers bypassed safeguards after the systems initially refused or blocked their requests.1
The report, requested by Sen. Maggie Hassan, became public Wednesday. Hassan said the findings show the threat “is no longer theoretical” and warned that without congressional action, a bad actor could use AI to plan and carry out a major attack. She plans to make AI safety a priority and hold hearings as she prepares to lead the Senate Homeland Security Committee.2
The companies acknowledged the risk of misuse but said the tests involved older models and that protections have since been improved. The GAO did not identify which chatbot produced which plan. It also stressed that its testing was narrow: each platform faced a single scenario, the sample was not statistically significant, and some details supplied by the chatbots were wrong. The watchdog made no recommendations.3
The findings arrived amid a separate debate about AI-assisted vulnerability hunting and a run of outages affecting banks, airlines and government systems. Intelligence Director Jay Clayton described the disruptions as maintenance-related; Sen. Mark Warner said a rush to apply patches could help explain them. Former cybersecurity agency director Chris Krebs said the outages were not attacks, but warned that vulnerability discovery and patching are far from over. The GAO report did not examine that issue, leaving the central concern unresolved: AI may help expose weaknesses, while also making harmful plans easier to assemble.4