The Washington Sun

Main Navigation

Subscribe Log in
  • Newsletters
    • Sign up for our newsletters
    • The Washington Sun newsletter
    • Afternoon Sun
    • Metro
    • Sign up for Dana Milbank’s newsletter
    • Sign up for Matt K. Lewis’ newsletter
  • Weather
  • Campaigns
    • 2026 Election
    • 2028 Election
    • Democrats
    • Republicans
    • Influence
  • Congress
    • House
    • Senate
  • Metro
    • Latest in Metro
    • Tom Sietsema
  • Sports
    • Latest in Sports
    • All-DMV High School
  • Policy
    • Administration
    • Defense
    • Energy & Environment
    • Immigration
    • Economy
    • Health
    • Foreign Policy
    • Housing
    • Technology
  • National
    • Legal
    • U.S. News
  • Perspectives
    • Latest Perspectives
    • Dana Milbank
    • Matt K. Lewis
    • Features
    • Forums
  • Podcast
    • Washington Sun Cast • Episode List
    • Milbank & Lewis • Episode List
    • Listen on Apple Podcasts
    • Listen on Spotify
    • Watch on Youtube
  • Who Are We?
    • About The Washington Sun
    • Events
    • Work for us
    • Press Releases
    • Support FAQ
  • Follow Us
    • X (Twitter)
    • Bluesky
    • LinkedIn
    • Instagram
    • TikTok
    • YouTube
Subscribe
Log in

Support independent journalism. Subscribe for $50 for your first year. Limited time.

Technology

Congressional Investigators Got Top AI Chatbots to Plan Terror Attacks

AI companies told the GAO their newer models are safer, but Sen. Maggie Hassan warns that older ones are still available.

ChatGPT logo seen on the App Store.

Democrats have pledged to uncover vulnerabilities and regulate the AI industry. (Matt Rourke/AP)

By Ayah Ali-Ahmad

November 18, 2026 05:00 a.m.

  • Email
  • LinkedIn
  • Twitter
  • Copy
    Link copied

Posing as regular users, congressional investigators got ChatGPT, Claude, Gemini and Grok to help construct plans for mass-casualty bombings and attacks on the power grid.

In less than 12 hours, each platform produced named targets, locations and methods, according to a Government Accountability Office report released Wednesday. One even generated a map with targets circled.

Two platforms laid out a vehicle bomb attack, one at a hotel parking garage and one at a convention center in major U.S. cities. Both picked a date and time. Both also offered tips on avoiding law enforcement, said Kevin Walsh, a GAO director who co-signed the report.

The other two chose electrical substations in a major U.S. city, one proposing a drone strike and the other gunfire. One estimated the outage would cause billions of dollars in damage.

Trending

  • U.S. Capitol building

    Bank CEOs, NSA Tell Congress That AI-Driven Patches Are Behind Outages

  • Police blocked off multiple blocks of 14th Street NW.

    What the Logan Circle Shooter Told Me

  • Trump China

    Xi Greets Trump on the Tarmac in China

  • Mary Peltola

    The Final Senate Race Is Called: Mary Peltola Flips Alaska Seat

The report, requested by Sen. Maggie Hassan (D-New Hampshire), sets the stage for how Democrats, soon to be in the majority, plan on investigating artificial intelligence companies’ products.

As lawmakers and the American public grow increasingly concerned about possible AI-driven doomsday scenarios, Democrats have pledged to uncover vulnerabilities and regulate the industry. Hassan, who is the top Democrat on the Joint Economic Committee and is expected to chair the Senate Homeland Security and Governmental Affairs Committee next year, is already building the case.

“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,” Hassan told reporters Tuesday.

Hassan said AI safety legislation will be “one of our first areas of legislative focus” once she takes over the committee. She said it would start with an early hearing on AI threats to homeland security and critical infrastructure.

The GAO randomly assigned each chatbot one of two scenarios: a mass-casualty bombing or an attack meant to cause economic damage. It was the first time the agency had done covert testing of AI platforms. The GAO tested the chatbots this spring.

According to the report, every AI platform refused requests to plan a violent attack at first, but investigators posing as average users — and paying for upgraded subscriptions on at least three platforms — successfully used “common safeguard-bypassing techniques,” which the report does not name.

Investigators concluded that the plans provided were “largely viable.” A bomb squad commander told the agency both bomb recipes “would result in an explosion,” and investigators found the required materials for sale at stores the chatbots suggested.

“To the companies’ credit, their safeguards have continued to evolve since we conducted these tests,” Walsh told reporters. “However, we conclude that, in its current state, this technology can be dangerous, and its misuse represents a growing risk to our nation.”

The chatbots’ plans did contain errors: One model selected a date when its target convention center had no scheduled event, a second generated a map with incorrect road placements, and some of the designated substations were behind high walls.

The GAO did not disclose its techniques or prompts, or say which chatbot produced which plan, to keep others from repeating the tests.

“The details behind the information provided by the AI platforms were not novel,” GAO investigators wrote. “Many of the specifics are available on the internet. However, the generation of plans that could have taken weeks or months were instead completed in hours with the assistance of AI.”

“To be clear,” Walsh said, “we are only able to say that in our single test of each of these companies, their safeguards failed.”

The companies acknowledged their chatbots could be misused but questioned how the tests were run.

“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,” Anthropic, which develops Claude, wrote in comments published with the report.

But Anthropic noted the GAO never compared the chatbots against what a web search would have turned up. “We ask that GAO either support such claims, strike them, or state plainly that it did not actually measure what uplift the models provided,” the company wrote.

The GAO agreed that kind of comparison was valuable but said 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.”

OpenAI 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.”

The GAO said that was the point. “This aligns with the actions of a malicious actor using an AI platform as a tool,” the agency wrote.

The companies also said the GAO tested models that have since been replaced.

Hassan warned that some of the older models are still available. At a minimum, Hassan said, Congress needs to pass “rules that prevent the release of models that don’t have adequate safety safeguards.”

“Whatever improvements they’ve made in new models don’t impact the models that GAO was able to access and circumvent the safeguards in,” Hassan said. She also called for a way to make sure safeguards are “updated on a regular basis” and to require companies to report illegal activity to law enforcement.

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.” The company noted the GAO used paid U.S. accounts. It said it “provides unique moderation policies for registered, paid US account holders,” which carry lower risk than free or foreign accounts.

Only one company suspended investigators’ accounts, according to the report. Two others removed parts of the chats within days — whether as a safeguard or by error, the GAO couldn’t tell — and one of those accounts was still active less than two weeks later.

Walsh declined to say whether any company alerted law enforcement. Account suspensions that came days or weeks later, he said, are “a trigger that may come too late.”

“We want the models to be safe. We want illegal activity to be reported,” Hassan said.

Author

Ayah Ali-Ahmad Ayah Ali-Ahmad is a reporter at The Washington Sun and an Allbritton Journalism Institute fellow.

The Latest

  • Jaland Lowe AP - 26318819711234

    Georgetown Men’s Basketball Is Undefeated and Unsatisfied

  • saltandsundry_111726_0081 copy.jpg

    How to Be a Good Host — or Guest — This Holiday Season

  • Zoo lights

    What to Do in the D.C. Area This Weekend (Nov. 19)

  • A member of the U.S. Secret Service stands watch at an event.

    Secret Service Faces Fresh Pressures as 2028 Preparations Begin

  • Wilson Building

    D.C. to Give $10 Million More to Support Struggling Ward 8 Hospital

Welcome back to The Washington Sun!

You can always update your profile and newsletters preferences on the My Account page.

Manage account settings

Welcome to
The Washington Sun

Subscribe to maintain unlimited access to independent journalism covering Washington's people, politics and power.

Continue to checkout ↗
  • About The Washington Sun
  • Work for us
  • Terms of Use
  • Subscription Agreement Terms and Conditions
  • Privacy Policy
  • Your CA Privacy Rights
  • Support FAQ
  • Contact us

© 2026.0 NOTUS MEDIA, LLC