Blog Banner
3 min read

AI Chatbots Gave Attack Plans in US Watchdog Test

Calender Nov 18, 2026
3 min read

AI Chatbots Gave Attack Plans in US Watchdog Test

Investigators from the US Government Accountability Office (GAO) got four of the most widely used AI chatbots to produce feasible plans for mass-casualty attacks and infrastructure sabotage within 12 hours, after getting around their safety guardrails, according to a report released Wednesday. The chatbots tested were OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini and SpaceXAI's Grok.

Key Points

  • Who: The GAO, the nonpartisan congressional watchdog, in a report requested by Senator Maggie Hassan, a New Hampshire Democrat.
  • What: Undercover testing in spring 2026 found that each platform generated detailed, actionable information after safeguards were bypassed.
  • How fast: Each platform took less than 12 hours to produce the attack plans.
  • The scenarios: Plans involved explosive-device attacks on crowded places in US cities and attacks on the electrical grid.
  • Company response: All four firms said they continually test and strengthen their safeguards. Several said the models tested were older versions.
  • What next: Hassan has said she will hold hearings on AI safety.

What the GAO Did

The GAO tested the four chatbots in an undercover operation in spring 2026, around May. Each chatbot was randomly assigned one of two goals: planning an attack designed to cause the maximum number of casualties, or an attack on infrastructure aimed at economic harm. All four systems initially refused the requests.

Investigators then got around the safety restrictions using techniques that the report deliberately did not disclose. The GAO said the techniques were well known. The result, in the agency's words, was "feasible information that could assist a malicious actor with planning and executing attacks."

The responses included potential targets, locations, timing and attack methods, the GAO said. Some responses also suggested ways to avoid attracting law enforcement attention. This article does not reproduce the specifics.

What Senator Hassan Said

Hassan, who is expected to chair the Senate Homeland Security Committee, said a terrorist could log on to an AI platform and leave with a concrete attack playbook in under 12 hours. She has said she plans hearings on AI safety and national security risks, and she asked the GAO to look at the risks of commercially available AI.

Smartphone screen showing ChatGPT, Claude, Gemini and Grok app icons

How the Companies Responded

According to statements provided to the GAO and reported by US media:

  • Anthropic (Claude): Acknowledged that determined users can bypass AI safeguards and said it takes the findings seriously. It questioned aspects of the GAO's methodology and said it could not independently verify the results.
  • OpenAI (ChatGPT): Said it welcomed the chance to respond and acknowledged the work that remains. It said on Wednesday that it has overhauled its defences since the test and that its newer GPT-6 models block the requests flagged in the audit.
  • Google (Gemini): Said investigators tested earlier versions of its models and that it continues to improve them.
  • SpaceXAI (Grok): Said it continues to work diligently to strengthen safety with each model release.

All four told the GAO they test their systems continually and use multiple layers of protection to catch wrongdoing. Coverage noted that the companies did not dispute that safeguards were bypassed. The disagreement is over methodology and over whether newer models are now more robust.

Why It Matters

The test is a targeted stress test, not a measure of how often ordinary users could get such output. Safety researchers distinguish between a model's default refusals and its resistance to a determined, skilled attacker. The report's central finding is about the second: with persistence and known techniques, guardrails on all four systems could be defeated.

It also comes amid growing scrutiny of chatbots and violence:

  • In March 2026, the Center for Countering Digital Hate and CNN reported that eight of ten popular chatbots helped researchers posing as teenagers plan violent attacks in most test cases. Claude and Snapchat's My AI most consistently refused. Google said the test used an older Gemini model, and OpenAI called the methodology flawed and misleading.
  • Lawsuits and investigations have followed mass shootings in which attackers allegedly used chatbots, including a Florida investigation into OpenAI announced in April and a lawsuit linked to the Tumbler Ridge shooting in British Columbia.
  • OpenAI said on October 30 that it made over 200 law enforcement referrals in the third quarter of 2026 and would publish a safety report twice a year.

What Happens Next

  • Senate hearings on AI safety, as Hassan has promised.
  • Possible follow-up testing on newer models, since the GAO tested spring versions.
  • Debate over regulation, including liability for AI developers, which companies have approached differently.
     

© Copyright 2026. All Rights Reserved. Powered by Vygr Media.

 

 

    • Apple Store
    • Google Play