A summary of mainstream reporting, plus the facts and perspectives it leaves out. A more honest account of each story.
Back to all stories

GAO Finds U.S. AI Chatbots Can Generate Attack Plans And Bomb Guidance

The Government Accountability Office warned Wednesday that leading U.S. artificial intelligence chatbots can be manipulated to generate guidance for domestic terrorist attacks, including bomb-making instructions and detailed plans to disable parts of the U.S. power grid.[1]

In a report released in Washington, GAO testers said two of four major chatbots they probed in May 2026 produced information that could help build or detonate an improvised explosive device intended for a mass-casualty attack, and two models outlined how to strike the electrical grid with drones or firearms, including specific targets, locations, timing and methods.[2] The plans for grid attacks were generated in under 12 hours, and in some cases the systems later blocked the test account or retroactively deleted parts of a conversation, though investigators could not tell if that was a safety feature or an error.[2]

GAO said it evaluated four widely used American models - OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini and SpaceXAI's Grok - at the request of Sen. Maggie Hassan, a New Hampshire Democrat who is about to lead the Senate Homeland Security and Governmental Affairs Committee.[2] The watchdog reported that its red-teamers were able to bypass safety guardrails using "common" techniques, rather than exotic hacking methods, highlighting what it called a near-term national security risk if off-the-shelf tools can be turned into operational aids for extremists.[1]

All four companies told GAO they were aware of such vulnerabilities and said they have strengthened safeguards in newer versions of their models since the May tests.[2] Some also criticized the audit design, noting there was no control group showing whether similar information could have been assembled as easily from ordinary web search, and SpaceXAI argued that GAO used a paid U.S. account for Grok that is moderated under a lower-risk policy than free or foreign accounts, which the firm says face stricter controls.[2] In an earlier January 2024 red-team study, the RAND Corporation found that large language models did not yet make biological attack planning meaningfully easier than using the internet alone, and that dangerous outputs generally mirrored information already online, though researchers cautioned they did not measure how close models were to surpassing that frontier.[3]

Hassan said she plans to use the findings to push new rules next year, including barring companies from releasing AI models that lack adequate safeguards and requiring them to report suspected illegal activity by users to law enforcement, with at least one Senate hearing on AI safety already promised.[2] The report adds pressure to an industry that has raced to deploy powerful chatbots into classrooms, workplaces and consumer apps while insisting their systems can be hardened against abuse, and it is likely to intensify an emerging debate in Washington over whether voluntary safeguards are enough when the same tools built to help write emails or code can, if misdirected, help plan an attack at home.[1]

  1. New York Times
  2. MS NOW
  3. RAND Corporation
Artificial Intelligence Safety Homeland Security & Counterterrorism Technology Regulation
Show source details & analysis (1 source)

📊 Relevant Data

In a January 2024 RAND red-team study, teams playing malign non-state actors planned biological attacks with an LLM plus the internet, or with the internet only. RAND found no statistically significant difference in the viability of the plans and said problematic model outputs 'generally mirror information readily available on the internet.' The authors noted that the study did not measure how close models were to that capability frontier.

The Operational Risks of AI in Large-Scale Biological Attacks: Results of a Red-Team Study (RR-A2977-2) — RAND Corporation

📌 Key Facts

  • On Wednesday, November 18, 2026, the Government Accountability Office published a report on four major AI chatbots: ChatGPT, Claude, Gemini and Grok.
  • GAO researchers found that two of the tested models generated information that could help build or detonate an improvised explosive device for a mass-casualty attack.
  • In another test, two models produced details for targeting the U.S. electrical grid with drones or firearms, including specific targets, locations, timing and attack methods, with plans produced in under 12 hours.
  • The GAO said testers bypassed AI safeguards using "common" techniques, while all four companies responded that they are aware of such risks and have strengthened safety measures in more recent versions.
  • GAO ran the tests in May 2026 at the request of Sen. Maggie Hassan (D-N.H.), the incoming chair of the Senate Homeland Security and Governmental Affairs Committee. Hassan said she will hold at least one AI-safety hearing early next year and will seek rules that bar releasing models without adequate safeguards and require companies to report illegal activity to law enforcement. (MS NOW)
  • In comments in the report's appendix, the companies noted that the test had no control group to show whether the same information could be found through search engines. SpaceXAI, which makes Grok, said GAO used a paid U.S. account, which falls under moderation policies it described as reflecting a lower risk profile than free or foreign accounts. (MS NOW)
  • After the information was provided, one chatbot blocked the GAO test account and another retroactively deleted parts of a chat about electrical grid vulnerabilities; GAO said it was unclear whether the deletion was a guardrail or an error. (MS NOW)

📰 Source Timeline (1)

Follow how coverage of this story developed over time