GAO Testers Got Four AI Chatbots to Draft Attack Plans

A 61-page GAO report says safeguards on ChatGPT, Claude, Gemini and Grok failed in covert tests, and the incoming Homeland Security chair wants hearings and new rules.

Key Points

  • GAO investigators got ChatGPT, Claude, Gemini and Grok to produce terrorist attack plans in less than 12 hours in covert tests.
  • Sen. Maggie Hassan, who requested the report, says she will hold a Senate Homeland Security hearing on AI risks in January.
  • The four companies said their models keep being refined, while Hassan warned older versions remain available.

WASHINGTON — Investigators posing as ordinary users got ChatGPT, Claude, Gemini and Grok to draw up plans for terrorist attacks in the United States, according to a Government Accountability Office report released on Wednesday.

Each platform produced a plan in less than 12 hours. The plans named targets, locations and timing, the report said, and set out methods for either a mass-casualty bombing or an attack on the electrical grid meant to cause economic harm.

Sen. Maggie Hassan, a New Hampshire Democrat who asked the GAO to do the work, said she will hold one of the Senate Homeland Security Committee’s first hearings in January on AI risks. Hassan is set to chair the committee when the new Congress takes over.

Hassan’s message on a call with reporters was blunt: “We need to act now.”

Her written statement went further. “The AI public safety threat is no longer theoretical,” Hassan 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.”

The GAO randomly gave each chatbot one of two jobs: plan a bombing meant to kill large numbers of people, or plan an attack meant to do economic damage. It was the first time the agency had tested AI platforms covertly. The audit ran from December 2025 to October 2026, and the covert testing itself happened in the spring.

For the bombing scenario, the platforms set out how to build and detonate an improvised explosive device. For the grid scenario, they turned to drones or firearms aimed at electrical substations. One answer even came with a map, its targets circled.

At first, all four said no. Investigators got past the safeguards anyway, using common bypass techniques. The report does not name those techniques. The GAO has not released its prompts, either, or said which platform produced which plan.

Outside experts checked what the chatbots produced. Investigators concluded the plans were largely viable, and a bomb squad commander told the agency the bomb designs in the tests would have resulted in an explosion. The report also noted errors. Roads on one map were in the wrong place. A few of the substations were tucked behind high walls, hard for an attacker to see. One platform settled on a date for its convention center target — a day when the center had nothing scheduled.

“The potential benefits of this technology are immense, but they can also benefit malicious actors,” the GAO wrote.

Kevin Walsh, a GAO director who co-signed the report, was careful about what the tests proved. “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 61-page report carries responses from Anthropic, Google, OpenAI and SpaceXAI, the companies behind the four platforms. Most said their models keep being refined.

Anthropic, which makes Claude, said it does not dispute that a determined tester can get harmful content out of an AI system by misrepresenting the purpose of a request, “and we take the finding seriously.” It said it could not verify the GAO’s findings.

Google said the testing was done on earlier models and that it continues to improve them. OpenAI said it welcomes the chance to respond, “as well as the work that remains.” SpaceXAI said it “continues to work diligently” to improve safety with each release.

Hassan said newer models do not settle the matter, because older versions remain available. She wants rules that stop models from being released without adequate safeguards. She also wants safeguards refreshed on a regular cycle, and a duty for companies to report illegal activity to law enforcement.

Comments
Advertisements
▶ Watch: Mountain Tang — Taste the Freshness
Follow Virginia Times on Facebook for news as it happensFollow
- Advertisement -
VT Newsroom
VT Newsroom
VT Newsroom is the editorial team of Virginia Times, published by Global News Break LLC. We cover Virginia, Washington and the world — politics, public policy, business, sports, entertainment and community news — reporting from primary sources and official statements whenever possible.

Latest news

Related news

LEAVE A REPLY

Please enter your comment!
Please enter your name here