GAO Report: Four Major AI Chatbots Generated Feasible Terror Attack Plans in Under 12 Hours

A terrorist could log onto an AI platform and walk away with a concrete attack playbook in under 12 hours.

“The AI public safety threat is no longer theoretical,” said Sen. Maggie Hassan, Democrat of New Hampshire, on Wednesday.

A new Government Accountability Office report, requested by Hassan and covered in detail by New York Post, found that four leading AI chatbots produced detailed, workable plans for terrorist attacks on US cities.

Investigators using Claude, Grok, ChatGPT and Gemini generated instructions for building and detonating an improvised explosive device, and for attacking the electrical grid with a drone and firearms — including specific dates and times for maximum carnage and where to park a bomb-laden car, New York Post reported.

The report says the instructions were “verified” and “feasible,” with law enforcement consulted to confirm the information. Hassan said a terrorist could log onto an AI platform and walk away with a concrete attack playbook in under 12 hours.

Two platforms produced mass-casualty improvised explosive device plans, while the other two outlined attacks on electrical infrastructure designed for economic damage, according to the GAO report itself. All four initially blocked the requests. All four were eventually circumvented using what the GAO calls “common safeguard-bypassing techniques,” which it declined to specify so the experiments cannot be recreated.

What the GAO actually found

The covert testing was conducted in spring 2026 on four of the most popular chatbots: OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini and SpaceXAI’s Grok. Each generated details such as potential targets, locations, timing and attack methods, the report said.

Government investigators find AI platforms produce terror plot plans in undercover op
Photo: nytimespost.com

“Each of these platforms provided detailed information that could be used to carry out attacks, such as specific targets, locations, timing, and attack methods,” the report said.

Researchers did not identify which models performed worse, citing fear the tests would be replicated.

The GAO’s own framing was blunt. “The potential benefits of this technology are immense, but they can also benefit malicious actors,” the report warned, citing growing risks to national and economic security. The watchdog also posted about the findings on Wednesday: “We tested AI platforms to see if they would create viable attack plans. They did.”

Where the companies push back

The companies told the GAO the tests ran on older model versions in the spring and said they have strengthened protections since.

Google said that investigators tested earlier versions of its models, saying, “We continue to improve our models.” OpenAI welcomed the opportunity to respond and acknowledged “the work that remains.” SpaceXAI said it “continues to work diligently” to strengthen safety with each release.

So the disagreement is narrow but real: the companies do not dispute that safeguards were bypassed. They dispute the currency of the evidence. Anthropic’s methodology objection is the sharpest — it accepted the underlying risk while contesting whether the GAO’s specific test results reflect its current models.

Ai chatbots provide chilling plans to plot mass casualty terror attacks bomb
Photo: wpnews.pro

Hassan Plans AI Security Hearing for January

Hassan, who is set to lead the Senate Homeland Security and Government Affairs Committee in the new year, said one of the panel’s first hearings would examine security risks posed by AI, including to critical infrastructure. She plans to prioritize legislation addressing the problem when Democrats control the chamber in January.

“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,” Hassan said in a statement Wednesday.

AI chatbots can help users plan out terror attacks, new study finds — WMUR-TV

What remains open: whether the spring testing on older model versions still reflects current behavior, which the companies contest and the GAO has not re-tested publicly. The full report, released Wednesday, stands as the first government-verified demonstration that consumer chatbot guardrails failed against ordinary bypass techniques on every platform tested.

Photo of author

Technology Desk

Technology Desk at Archyde.

Seattle Doctor Uses LinkedIn to Overturn Insurance Denials for Patients

Sandra Kaufmann Solos at Loyola Dance Program’s 20th Anniversary Concert