GAO: chatbots gave attack plans
Partly confirmed
Importance: major (4 of 5)The takeaway
Hassan plans hearings early next year; the companies disputed parts of the method and said safety has improved since May.
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- Major (4 of 5)
- Last verified
- 20 November 2026
Your AI and this story
- GPT-6.2 Astra171 days after its cutoff
- Claude Opus 5.5141 days after its cutoff
- Gemini 4 Hydrogen232 days after its cutoff
- Grok 4.7171 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 141 days before it.
Key facts
- Tests run in May 2026 at Sen. Hassan’s request; report released 18 Nov; it does not attribute responses to individual chatbots (MS NOW)
- At least one chatbot blocked a GAO account after the information had already been given
- Hassan: ‘If Congress doesn’t act, it is a question of when, not if, a bad actor will use AI to plan and carry out a major attack’; she expects to chair Senate Homeland Security and Governmental Affairs and wants rules barring releases without adequate safety protocols, reporting of illegal activity to law enforcement, and regular safeguard updates
- Companies: no control group comparing with search engines; improvements since May; GAO used paid accounts; SpaceXAI said its moderation differs for paid US accounts
- GAO report: ‘Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans’ (GAO-27-108814); per summaries, two platforms gave information for building and detonating a mass-casualty explosive device and two others grid-attack plans involving a drone and firearms, and each case did trigger safeguards (refusals, removals, account suspensions)
- Second wave (Nov 18-20): NYT (Dustin Volz), ‘American A.I. models have potential to be turned against U.S., study finds’, reads it as a sign Democrats will push AI rules in the new Congress; CBS: ‘Senate AI hearings in works’; USA Today, FedScoop, NewsNation, Tech Times and the NY Post covered it; Futurism separately reported that Claude can be tricked into choosing drone strikes when told it is a game (the Vals study)
What happened
Congress’s own auditor tested the four best-known US consumer chatbots and reported that their refusals could be bypassed. The report is the first GAO-level official test of this kind; the specific prompts and outputs are not reproduced here.
Sources
4 sources from 4 sites. Numbers match the chips in the text.
4 sources: 1 primary, 3 press
Primary
- GAO: Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plansgao.gov, official
Press
Changes
- GAO report link and second-wave coverage (news agent)
- Filed (GAO report itself not yet located; facts from MS NOW, NYT and Dayton Daily News headlines)