AI models used to plan 'viable' attack plots in test despite safeguards: Government report

Sen. Maggie Hassan warns that AI could be misused by bad actors.

The world's leading artificial intelligence platforms were used to create "viable" plans to build homemade bombs and carry out mass casualty attacks in major U.S. cities in a test, according to a new government report released Wednesday.  

Using Claude, Grok, ChatGPT and Gemini, investigators with the nonpartisan Government Accountability Office produced "feasible information that could assist a malicious actor with planning and executing attacks" within 12 hours, using well-known techniques to bypass the platforms' guardrails, the agency wrote in its report.

The report from the congressional watchdog comes at a time when breakthrough advances in AI have underscored the promise and potential of the rapidly evolving technology and raised new concerns by lawmakers in Washington about the potential for misuse.

Close up of hands typing on a computer keyboard.
D3sign/Getty Images

In recent years, gunmen have used A.I. to plan mass shootings in Florida and Canada, according to authorities. And one top A.I. company disclosed in September that users in Yemen unsuccessfully attempted to use its model to build missiles. 

"The same information that testers were able to get ahold of could be deadly in the hands of a terrorist who actually intends to carry out an attack," said Sen. Maggie Hassan, D-N.H., who requested the study. 

Last April, researchers randomly assigned each of the platforms to one of two scenarios: an improvised explosive device (IED) attack, and an attack on the U.S. electrical grid.

Two of the AI platforms generated distinct plots to build, deploy and execute IED attacks in major cities. They suggested specific dates, times and targets -- a hotel parking garage and a convention center -- around scheduled public events.  

PHOTO: Senate Treasury
Sen. Maggie Hassan, D-N.H., asks questions following Treasury Secretary Scott Bessent's testimony before the Senate Committee on Finance hearing to examine the President's proposed budget request for fiscal year 2027, Wednesday, June 3, 2026 in Washington.
Allison Robbert/AP Photo

Investigators ran the experiment with each platform once, but randomized the tests and did not disclose which platforms produced each result.

Both platforms suggested ways to avoid the attention of law enforcement; one produced a timeline for planning and executing the attack that recommended creating a shell company and purchasing supplies across multiple states.

"At minimum [one] platform provided enough information that could help a malicious actor to avoid detection and produce a very large IED," the agency wrote in its report.

Investigators verified the accuracy of the platforms' instructions with surveillance, research and by consulting a bomb squad commander at a "major metropolitan area" fire department.  

The other two platforms provided instructions for targeting electrical grid substations in major U.S. cities, identifying specific drone and firearm models and how to acquire them.

They also generated a map and image to help visualize the execution of the planned attacks.

In statements to the GAO, Anthropic (Claude), Google (Gemini), OpenAI (Chat GPT) and SpaceXAI (Grok), told the agency that they continually test their systems and update and expand safeguards built into their platforms, and have multiple layers of security to catch wrongdoing.

All have taken steps since the study was conducted last spring to improve their models and safeguards, the companies told the GAO.

"The companies also assure us that this is an area of concern and ongoing research," said Kevin Walsh, the director of IT and cybersecurity audits at the GAO, 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 agency, which redacted specific results and model information from their tests and report to prevent misuse of their work, shared their findings with the artificial intelligence companies upon request.

The GAO also did not specify in its report if any of their experiments prompted a response from law enforcement but noted that some of the platforms suspended their accounts after the study was conducted.

Hassan, the incoming chair of the Senate Homeland and Governmental Affairs Committee, said she plans to hold hearings on AI safety next year and work on new legislation to beef up safety requirements.

"These companies are facing extraordinary competitive pressures to get their models out the door and to the public as quickly as possible," she said.

"One of the things that will be necessary going forward to make sure that they're all dealing with the level playing field when it comes to safety, is making sure that before models are released, they have safeguards that can't be breached in this way."

In a statement to ABC News, an OpenAI spokesperson said: "We take these findings seriously and appreciate the opportunity to work constructively with GAO, Senator Hassan and the relevant committees to address them. While the study findings are concerning, the testing took place months ago on an older version of ChatGPT, and we've made considerable safety improvements since then, including with our GPT-6 family of models. We rigorously test our safeguards internally and with external experts, and we use monitoring, human review and enforcement to identify and address misuse and strengthen our protections."

ABC also reached out to Anthropic, Google and SpaceXAI.

Sponsored Content by Taboola