Government watchdog gets ChatGPT, Claude, Gemini and Grok to draft terror attack plans
Sen. Maggie Hassan, who requested the test, plans AI hearings in January, while the companies say they have strengthened safeguards since spring.
The Government Accountability Office, a congressional watchdog, reported on Wednesday that four of the most-used AI chatbots produced plans for terrorist attacks when its investigators tested them in spring 2026. The chatbots were ChatGPT, Claude, Gemini and Grok. The test was requested by Sen. Maggie Hassan, a New Hampshire Democrat who is due to chair the Committee.
Each chatbot first refused, but investigators got around the safeguards. Each platform took less than 12 hours to produce its plans. Two chatbots drew up plans for improvised explosive devices meant to cause mass casualties, and two drew up plans to attack the electrical grid to cause economic harm.
The plans included details such as targets, locations, timing and attack methods. The GAO consulted experts to check how feasible they were. The companies behind the chatbots, Anthropic, Google, OpenAI and SpaceXAI, said they keep improving their safety measures.
Google said the testers used earlier models, and Anthropic said it could not verify the findings. Hassan said the AI public safety threat is "no longer theoretical." She said one of her first hearings when the new convenes in January will focus on AI risks.
- less than 12 hourstime each platform took to generate its attack plans, per the GAO reportThe Boston Globe
- 61-pagelength of the GAO reportThe Boston Globe
Three sides of the story
13 news outlets reported this
Tap a color or swipe sideways to switch sides.
Liberal7 outlets
Hassan's alarm and the push for AI safety rules
The Independent calls the test "alarming" in its headline and says it raised "serious alarm bells." It ties the report to outages that affected air traffic control and online banking, and to Homeland Security Secretary Markwayne Mullin, who is set to appear before a House committee on Wednesday. The Boston Globe runs an Associated Press report that opens on Hassan vowing hearings.
It adds the companies' letters, including Anthropic saying it takes the finding seriously but cannot verify it. MS NOW goes into the most detail. It describes a hotel parking garage bomb plan, a rental van and a drone diagram, and notes the companies' points that the test had no control group and that SpaceXAI says GAO used a paid account.
It reports Hassan's wish for rules blocking release of models without adequate safeguards. The New York Times headline says American AI models "have potential to be turned against U.S." The Independent, the Globe and MS NOW quote Hassan warning that if Congress does not act, a bad actor's attack is a question of when, not if.
- The Independent: Congressional watchdog gets AI chatbots to plan terror attacks against America in alarming test
- The Boston Globe: Senator Maggie Hassan vows AI hearings after watchdog gets platforms to plan terrorist attacks
- MS NOW: Government regulators asked AI chatbots to plan a terrorist attack. The chatbots did.
Center2 outlets
Outlets on this side gave the story little coverage.
Conservative4 outlets
Investigators' verified, feasible terror plans from chatbots
Fox News leads with investigators who "successfully bypassed" safeguards on four platforms and got plans in under 12 hours. It reports that the GAO consulted experts and made real-world checks, and that the GAO found the information feasible enough to help someone planning an attack. It gives the companies' replies at length, including Anthropic questioning aspects of the GAO's methodology and Google's "We continue to improve our models."
The New York Post calls the plans "chilling" in its headline and stresses "strategic bomb placement." It says the report called the instructions "verified" and "feasible," after investigators consulted law enforcement. It quotes Hassan that a terrorist could get "a concrete playbook" in under 12 hours.
It closes with the companies telling the GAO that the tests used older versions and that protections have since been strengthened.
Which side made the stronger case?
We will email you when the story continues. At most one email a day.
All sources: 13 articles from 13 outlets
Liberal (7)
- Business Insider: A government watchdog asked AI chatbots to plan bombings and attacks on US power grids. It found that some would've worked.
- CBS News: AI platforms planned U.S. terror attack when asked to
- MS NOW: Government regulators asked AI chatbots to plan a terrorist attack. The chatbots did.
- The Boston Globe: Senator Maggie Hassan vows AI hearings after watchdog gets platforms to plan terrorist attacks
- The Independent: Congressional watchdog gets AI chatbots to plan terror attacks against America in alarming test
- Yahoo News: Report reveals that AI models could plan 'viable' terror attacks despite safeguards
Center (2)
- Forbes: ChatGPT, Claude, Gemini And Grok Produced Viable Terror Attack Plans, Government Finds
- NewsNation: AI chatbots capable of planning theoretical terror attacks: Report
Conservative (4)
- Fox News: Government investigators find AI platforms produce terror plot plans in undercover op
- New York Post: AI chatbots provide chilling plans to plot mass-casualty terror attacks — including strategic bomb placement
- The Epoch Times: Congressional Watchdog Got AI Platforms to Plan Terror Attacks in Simulation
- The Washington Times: Claude, ChatGPT, Gemini and Grok produced attack plans in covert tests: GAO report
The evening briefing by email
One email a day, around 8:30 p.m. New York time: the seven main stories, each with its three sides, and the blind spot of the day. Unsubscribe with one tap.
A Telegram channel is coming soon.