✉ The Friday AI Brief: the week's 5 best AI stories, tools & comparisons — in your inbox every Friday morning.

Seal of the United States Government Accountability Office (GAO)

Government Watchdog Got ChatGPT, Claude, Gemini and Grok to Plan Terror Attacks in Under 12 Hours

Washington, November 18, 2026 — Government investigators posing as ordinary users got all four of America’s most-used AI chatbots — ChatGPT, Claude, Gemini and Grok — to produce detailed, viable plans for terrorist attacks, according to a 61-page Government Accountability Office (GAO) report released Wednesday. Each platform produced its plan in less than 12 hours. Why it matters: this is the first covert government test of commercial chatbots, and the senator who ordered it is about to chair the Senate Homeland Security Committee — she says hearings and AI safety legislation come first in January.

What happened

The GAO randomly assigned each chatbot one of two scenarios: a mass-casualty bombing, or an attack on the electrical grid meant to cause economic damage. Two platforms produced plans for constructing and detonating an improvised explosive device; the other two produced plans to attack electrical substations, one proposing a drone strike and the other gunfire. The plans included specific targets, locations, dates and times, and attack methods — one platform even generated a map with targets circled, and both bombing plans offered tips on avoiding law enforcement.

Every platform refused at first. Investigators — who paid for upgraded subscriptions on at least three platforms — then got through using what the report calls “common safeguard-bypassing techniques,” which it does not name. The GAO also withheld which chatbot produced which plan, to keep others from repeating the tests.

How solid were the plans?

Investigators concluded the plans were “largely viable.” A bomb squad commander told the agency both bomb recipes “would result in an explosion,” and investigators found the suggested materials for sale at stores the chatbots named. The plans were not flawless: one model picked a date when its target convention center had no event scheduled, another drew a map with roads in the wrong places, and some chosen substations sat behind high walls.

The GAO was candid about the limits of its own test. “We are only able to say that in our single test of each of these companies, their safeguards failed,” GAO director Kevin Walsh told reporters. The agency also noted the underlying details were not novel — much of the information exists on the internet — but that work which could have taken weeks or months was completed in hours with AI assistance.

What the companies said

The testing happened this spring, on models the companies say have since been replaced. Anthropic wrote that it does not dispute that a determined tester can get harmful content out of an AI system by misrepresenting a request, and that it takes the finding seriously — but it also challenged the GAO to measure how much “uplift” the models actually provided over an ordinary web search. OpenAI said it reran all 325 reviewed requests on its newer GPT-6 Astra model, which “generally provided more refusals and safer alternative responses.” SpaceXSI, which makes Grok, said successful jailbreaks require “adversarial prompting.” Only one company suspended the investigators’ accounts during the tests.

Sen. Maggie Hassan (D-N.H.), who requested the report, warned that some of the older models tested are still available to users. “The AI public safety threat is no longer theoretical,” she said. “If Congress does not act, it is a question of when — not if — a bad actor will use AI to plan and carry out a major attack.” She said one of her committee’s first hearings in January will focus on AI risks, and she called for rules preventing the release of models without adequate safeguards, plus a requirement to report illegal activity to law enforcement.

Why it matters

For anyone who uses these tools, the practical takeaway is not that chatbots are bomb manuals — the GAO itself says the raw information was already online. It is that the safeguard layer, the thing standing between a normal account and the worst requests, failed in every single test, on every major platform, within half a day. That is now a documented government finding rather than an activist claim, and it lands just as Congress prepares to debate mandatory AI safety rules. Expect the companies’ “newer models are safer” argument to be tested — publicly, under oath, in January.

Frequently asked questions

Which AI chatbots did the GAO test?

ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google) and Grok (SpaceXSI) — described by the GAO as the four most-used platforms. The report does not say which platform produced which attack plan.

Did the chatbots refuse the requests?

Yes, initially. Investigators then bypassed the safeguards using common techniques the report does not disclose, and each platform produced a plan in less than 12 hours.

Are the tested models still available?

The companies say the spring tests ran on older models that have since been replaced with safer versions. Sen. Hassan warned that some of those older models remain accessible.

What happens next?

Hassan, set to chair the Senate Homeland Security and Governmental Affairs Committee in January, says an early hearing will examine AI threats to homeland security and critical infrastructure, and that AI safety legislation will be a first legislative focus.

Sources: Associated Press; The Washington Sun; New York Post; Government Accountability Office report coverage, November 18, 2026.

Leave a Comment

Your email address will not be published. Required fields are marked *

Get the 5 best AI tools every week

Top AI news, tools, and prompts — one short email. Free, unsubscribe anytime.

Run a newsletter of your own? Monetize and grow it with SparkLoop →

Scroll to Top