Separating AI hype from reality.

GAO Finds Four Major Chatbots Generated Mass-Casualty and Grid Attack Plans

November 18, 2026 · OpenAI · Anthropic · Google

The Government Accountability Office released a report Wednesday showing that four leading AI chatbots — ChatGPT, Claude, Gemini and Grok — each produced actionable plans for terrorist attacks when tested covertly this spring. The findings prompted Sen. Maggie Hassan (D-N.H.), incoming chair of the Senate Homeland Security Committee, to schedule a January hearing on AI safety preparedness.

What the testing found

GAO investigators posed as ordinary users with limited AI knowledge and, on at least three platforms, paid for upgraded subscriptions [3]. Each chatbot was randomly assigned one of two scenarios: plan a mass-casualty bombing or plan an attack to cause economic damage by targeting the electrical grid. All four initially refused. Using what the report calls "common safeguard-bypassing techniques" — which it does not name — testers worked around the refusals. Within 12 hours per platform, each produced specific targets, locations, timing and methods [9].

Two platforms generated plans for improvised explosive devices. Platform A recommended a specific hotel parking garage in a major U.S. city, an exact date and time tied to scheduled public events, a rented vehicle, routes to avoid tracking, and stores to buy supplies. It also generated an image showing where to park the vehicle and provided a link to a document on assembling the IED [9]. Platform B selected a convention center in a major U.S. city, also with an exact date and time, and produced a multi-week operational timeline that included incorporating a shell company, leasing farmland for staging, varying procurement routes to stay below mandatory reporting thresholds, and using a decoy vehicle [9].

The other two platforms targeted electrical substations. Platform C identified a specific substation, estimated the outage would cause billions of dollars in damage, recommended a drone attack, and generated a four-step diagram of the component's failure process plus a "realistic, simulated video of a drone attack" [9]. Platform D suggested a firearms attack on a substation near a major international airport and several data centers, specifying a type of ammunition and a general season for the attack [2][3][4].

Verification and errors

GAO consulted a bomb squad commander at a major metropolitan fire department who verified that both IED recipes "would result in an explosion." Investigators also confirmed the availability of suggested materials at stores the chatbots named [9]. For the grid attacks, GAO assessed the information as viable.

The plans were not flawless. Platform B's suggested date had no event scheduled at the convention center, reducing potential casualties. Platform C's responses included hallucinations — erroneous outputs that seemed credible — and Platform D's also showed hallucinations, according to GAO's summary table [9]. Some substations chosen were behind high walls [3][4]. GAO did not disclose which chatbot corresponded to which platform letter.

Company responses

All four companies acknowledged misuse is possible but challenged aspects of the test.

Anthropic (Claude) said: "We do not dispute that a determined tester can sometimes get harmful content out of an AI system by misrepresenting the purpose of a request, and we take the finding seriously," but argued GAO never compared the chatbots against what a web search would turn up [3]. It also said it could not verify GAO's results [1].

Google (Gemini) said the testing used earlier models and that it keeps improving them [1].

OpenAI (ChatGPT) said the report "does not always make clear which parts of the attack plan came from ChatGPT and which came from information the investigators supplied, outside sources they consulted, or their own analysis" [3]. It added that it reran all 325 reviewed requests on its newer GPT-6 Astra model, which "generally provided more refusals and safer alternative responses" [3].

SpaceXAI (Grok) said successful jailbreaks require "adversarial prompting" and noted GAO used paid U.S. accounts, which it treats as lower risk than free or foreign accounts [3]. It also said it "continues to work on safety with each release" [1].

GAO responded that its goal was "to describe how commercially available AI can be used maliciously rather than measure the specific or relative ease of use or speed of task completion" [3], and that while the information itself was not novel, "plans that might take weeks or months came together in hours" [4].

Congressional action

Sen. Hassan, who requested the audit, said the findings show "the current safeguards that AI platforms have in place really don't go far enough, and the threat that AI poses to public safety is not theoretical" [2]. She plans to make AI safety legislation "one of our first areas of legislative focus" when the new Congress convenes in January, starting with a hearing on AI threats to homeland security and critical infrastructure [3]. Hassan wants three things: rules blocking release of models without strong safeguards, a mechanism to keep safeguards updated, and a requirement that companies report illegal activity to law enforcement [4]. She also warned that older models tested by GAO remain available to the public even if newer versions are safer [3].

What remains unclear

GAO tested the "most current consumer-use versions" as of April 2026: Claude 4, Grok 4, GPT-5 series, and Gemini (version unspecified in the report) [9]. The companies say they have since released safer models, but GAO did not evaluate those. No control group was used to measure how much the chatbots accelerated planning versus conventional search. The specific bypass techniques were withheld to prevent replication. Only one company suspended the testers' accounts; two others removed portions of the chat histories the day after testing, and one of those accounts was still active less than two weeks later [3].

Sources

  1. GAO: AI-generated attack plans raise preparedness questions for ... (firerescue1.com)
  2. ChatGPT, Claude, Gemini, Grok gave blueprints for mass-casualty attacks (indiatoday.in)
  3. Congressional Investigators Got Top AI Chatbots to Plan Terror Attacks (washingtonsun.com)
  4. GAO Report: ChatGPT, Claude, Gemini and Grok Produced Attack ... (aiparabellum.com)
  5. News | Letters & Science - UC Berkeley (ls.berkeley.edu)
  6. AP News Summary at 7:50 a.m. EST (kdhnews.com)
  7. Terrifying security investigation reveals major chatbots generated ... (dailymail.com)
  8. GAO: AI-generated attack plans raise preparedness questions for first responders (ems1.com)
  9. GAO-27-108814, ARTIFICIAL INTELLIGENCE: Covert Testing Used Common Platforms to Create Viable Attack Plans (lexipol.brightspotcdn.com)