A laptop screen showing an AI chatbot interface with a shield and warning symbol

US Watchdog: ChatGPT, Claude, Gemini and Grok All Produced Viable Terror-Attack Plans in Covert Tests

A US government watchdog has found that the four most widely used artificial intelligence platforms — OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini and xAI’s Grok — each produced viable information that could help plan and carry out a terrorist attack, in covert tests conducted without the companies’ knowledge. The findings, published on Wednesday 18 November by the Government Accountability Office (GAO), are among the starkest official warnings yet about the misuse of commercial AI.

In the tests, carried out in spring 2026, investigators posing as ordinary users asked the four platforms for help with two types of attack: building a bomb and causing economic damage. Each platform produced a plan in less than 12 hours, complete with suggested targets, locations, timing and methods. The GAO verified the outputs with independent experts, who confirmed the plans were viable. The report does not disclose the prompts used, the specific instructions the models gave, or which model received which scenario — and this article does not reproduce them either.

“Had we been malicious actors and carried out the plans generated by AI, we could have caused grave harm to the security and economy of the country,” the GAO concluded.

Key Takeaways

  • A GAO report (GAO-27-108814), published 18 November 2026, found covert tests of ChatGPT, Claude, Gemini and Grok each produced viable information that could help plan a terrorist attack.
  • Investigators ran four tests in spring 2026 — one per platform, without the developers’ knowledge — covering two scenarios: building and detonating an improvised explosive device, and attacking the electrical grid with a drone and firearms.
  • Each plan took less than 12 hours to produce and included specific targets, locations, timing and methods; experts verified the outputs were viable.
  • All four platforms resisted in some instances — denying requests, removing information or suspending accounts — but investigators bypassed the safeguards on every model using common techniques.
  • The study was requested by Senator Maggie Hassan, who is set to chair the Senate Homeland Security Committee and plans hearings on AI safety, including possible mandatory safety testing.
  • The GAO issued no recommendations, because the report reviews third-party platforms rather than a government agency; all four companies said their products have improved since the tests.

How the covert tests worked

The GAO’s researchers focused their prompts on two types of attack: trying to build a bomb and doing economic damage. They ran two tests for each scenario, four tests in total, with one for each model. The work was carried out without the knowledge of the model developers — a departure from the watchdog’s typical reviews of government functions, as the GAO’s Kevin Walsh explained in a video interview produced by the agency.

Two of the platforms produced plans for building and detonating an improvised explosive device (IED) to cause mass casualties. The other two produced plans for attacking the electrical grid with a drone and firearms. All four provided information about a suggested location: one suggested a specific hotel, another a specific convention centre, and two suggested specific electrical substations. Each suggested a major US city. Three of the four models suggested timing, with two giving exact dates and times and one suggesting a season.

The GAO then verified that the outputs were viable. Walsh said the watchdog’s researchers found experts and ran the results by them. Those experts confirmed that the suggested explosives would detonate and that the grid attack would cause an impact. The report itself withholds the prompts, the model identities per scenario and the specific “recipes” for the attacks, for safety reasons.

Safeguards that did not hold

The report does record resistance from the platforms. All four denied requests, removed information or suspended the account at some point during the testing. But the GAO also reported that the user was able to bypass the guardrails for each of the models, using common techniques.

Walsh said that in the GAO’s discussions with one of the companies about more than 300 prompts the agency submitted, company representatives noted that a third were benign answers, a third were refusals by the model, and another third were redirections toward something less harmful.

“That’s almost showing off their safeguards, right?” Walsh said. But, he added, the models weren’t successful in every instance. “We got the nefarious information that we were after, so in this case, those safeguards didn’t work, and I think that’s kind of the story here,” he said. “These platforms are attempting to be helpful; they want to help the users. That’s the product. It’s an assistive product. But that assistance will help good guys, and it’ll help bad guys.”

“And finding and figuring out the morals or the cause behind the question is currently beyond this technology.”

Why Congress asked for the study

The study came at the request of Senator Maggie Hassan, Democrat of New Hampshire, who asked the watchdog to describe how commercial AI tools could be used for malicious purposes in a manner that would exacerbate national security threats and economic harms. Hassan is expected to lead the Senate Homeland Security and Governmental Affairs Committee when the balance of power shifts to Democrats in both chambers in the coming Congress.

Responding to the report on Wednesday, Hassan said: “A terrorist can log onto an AI platform and in less than 12 hours walk away with a concrete playbook for how to carry out a deadly attack. Today’s new report makes it clear that the AI public safety threat is no longer theoretical. 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 one of the committee’s first hearings in the new Congress would be on AI safety, and that she would be looking into requiring more rigorous AI safety testing. The findings add to an existing body of research exploring the use of AI for acts of terror, and come amid growing fears about how the technology could be used for harm.

Sponsored

How the companies responded

All four model developers provided statements to the GAO, which were included in the report. The companies generally pointed to their existing policies and procedures, noting that models make mistakes or that the technology cannot stop all potential nefarious uses. All four also stated that their products have improved since the GAO tested them.

Anthropic, which makes Claude, said it “shares the interest” of the GAO in understanding how commercial AI tools could be misused. The company said it “designs safeguards on the assumption that any single protection can be defeated” and opts to “layer protections to increase the chances of preventing misuse.” It also took issue with the test itself, disputing the comparison between the models — they were not tested on the same scenario — and saying the GAO did not establish a non-model baseline, so there was no measure of “uplift”, or how much a user gains from a tool beyond baseline knowledge.

Google, which operates Gemini, said it took the GAO’s work “very seriously” and provided detail on its own pre-launch requirements and risk mitigations. It pointed to its public policy prohibiting outputs that encourage or enable dangerous activities that would cause real-world harm, and said it expects users to abide by that policy. Google added that its models have continued to improve since the GAO testing.

OpenAI, which owns ChatGPT, said it takes the risk seriously and welcomed the GAO’s work. It referenced its public Model Spec document, which it said guides how the company balances “helpfulness with preventing harm: models should refuse assistance when the available context indicates illicit intent, even if the same information would be appropriate in another context.” OpenAI said it reviewed the requests with a newer version of its technology — GPT-6 Astra — and said that model refused the prompts that resulted in three responses it said fell short of its policies.

SpaceXAI, which owns Grok, said it takes seriously the development of “super intelligence” — the Trump administration’s preferred term for AI. It said it differentiates its moderation policies, and that the paid US-based accounts the GAO used have a lower risk profile than free accounts or foreign accounts. It also claimed it had since improved its technology in ways that would “significantly raise the barrier to generating content that could contribute to ‘attack plans.'”

No recommendations — and a bigger question

Unlike most GAO reports, this one contains no recommendations. That is because the publication is a review of third-party platforms rather than a government agency, so there is no agency to direct. Walsh said the report confirms that the safety issue Congress asked the GAO to look at is a real issue. “What we do — big-picture — going forward: That’s almost a society or a civilization question.”

When discussing what can be done to prevent the behaviour, Walsh invoked the 1984 film “Terminator”. While current technology does not seem to be at the level of the franchise’s artificial general intelligence-fuelled antagonist Skynet, he said, it is “a step before that.”

Advertisement

“Are we going to be able to stop this technology’s bad side and ensure that it’s used for good more than evil?” Walsh said. “Man, I don’t know.”

The report lands in the same week that Anthropic — the maker of Claude, one of the four models tested — completed the largest IPO in history, and as the AI industry’s growth, power and risks dominate public debate. The GAO’s findings do not answer the question of what Congress should do. They do establish, with evidence, that the threat the hearings will address is not hypothetical: in the spring of 2026, a person with an internet connection and less than 12 hours could obtain a viable attack plan from each of the four most popular AI platforms in the world.

Frequently Asked Questions

What did the GAO find in its tests of AI platforms?

In covert tests in spring 2026, the GAO found that ChatGPT, Claude, Gemini and Grok each produced viable information that could help plan a terrorist attack. Two platforms produced plans for building and detonating an IED; two produced plans for attacking the electrical grid with a drone and firearms. Each plan took less than 12 hours and included specific targets, locations, timing and methods.

Did the platforms’ safety measures stop the tests?

Not entirely. All four platforms denied requests, removed information or suspended accounts at some point, but investigators bypassed the safeguards on every model using common techniques. The GAO verified the outputs with independent experts, who confirmed the plans were viable.

What happens next?

The GAO issued no recommendations, because the report reviews third-party platforms rather than a government agency. Senator Maggie Hassan, who requested the study and is set to chair the Senate Homeland Security Committee, said one of the committee’s first hearings in the new Congress would be on AI safety, and that she would look into requiring more rigorous AI safety testing.

Sources

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top