Latest

Artificial Intelligence

GAO gets four AIs to plan an attack in the US

A report by the US regulator obtained viable attack plans from ChatGPT, Claude, Gemini and Grok in under 12 hours. The Senate is preparing hearings.

Translated from Spanish by an automated system. The Spanish original prevails. Read the Spanish original

GAO gets four AIs to plan an attack in the US
GAO gets four AIs to plan an attack in the US

Federal auditors sat down in front of a chatbot, asked it for an attack plan and received, in under twelve hours, a proposal concrete enough to attempt. They repeated it with the four most widely used AI platforms in the United States, with the same result every time. The GAO report on the attack, which the agency published on October 16 and circulated on November 18, documents an undercover test with ChatGPT, Claude, Gemini and Grok: all four activated their safeguards and all four were bypassed using known techniques.

What the document measures is not the existence of dangerous material, which has been on the internet for years: it is the time needed to turn it into a plan, and AI compressed that to hours.

Politics moved the same day. Senator Maggie Hassan, who will chair the Senate Homeland Security Committee from January, announced hearings on AI safety and said she wants rules to prevent the release of models without sufficient safeguards. AP adds a detail that gives weight to the hearing: the investigation originated in Hassan's office, so the report is the starting material for those sessions.

Two scenarios, four letters and a single pass

The work began in December 2025 and closed in October 2026, with the tests concentrated in the spring. The auditors chose two scenarios with the help of experts and settled on Claude, Grok, ChatGPT and Gemini, the four most widely used platforms in April 2026 and chosen on volume alone. The rest of the design was intended to ensure no one could point to a company by name.

Each platform randomly received one of the two scenarios and a letter instead of its name, so the report does not say which company produced which plan. GAO ran the test only once per platform and warns in the text that the result is a snapshot of a moment, not a statistically significant sample.

There were two scenarios: an attack with an improvised explosive device in a vehicle, designed to cause mass casualties, and an offensive against the power grid with drones and firearms, aimed at economic damage.

To check whether what they received was real, the researchers carried out searches, surveillance and interviews with specialists. Among those consulted was an explosive ordnance disposal squad leader from a major city fire department, someone who knows better than anyone whether a bomb plan holds up on paper.

The four results, letter by letter

Four rows and a yes in the two columns that matter. The table answers two questions per platform: whether the information was enough to plan and execute an attack and whether the user managed to bypass the safeguards. In all four cases the answer is the same. The letters are those used by GAO and do not allow the company to be identified, because the document deliberately conceals that correspondence.

Letter Assigned scenario Precision of the plan
A Explosive device in a vehicle Exact date and time
B Explosive device in a vehicle Exact date and time
C Drone attack No time specification
D Firearms attack Season of the year

Two of the four plans arrived with an exact date and time, and GAO deemed them viable. The other two stopped at the season of the year or at no time specification at all, and passed the same filter. Temporal precision was therefore not what separated one platform from another.

In all four cases the platform activated its defenses, either with a refusal or by withholding the information, and in every case the researcher found a way to continue. That identical pattern across four products from four different companies is the finding that none of them disputes in their responses.

ABC News picks up a detail that does not appear in the public summary: two platforms added on their own advice on how to avoid attracting the attention of the authorities, something no one had asked them for. The report does not reproduce that content or the evasion techniques used by the researchers, removed from the document so they cannot serve as a recipe.

The question of consequences remains. GAO writes that it has no knowledge of whether any of its tests triggered a law enforcement response: four viable plans passed through four commercial systems and the document records no effect outside the exercise.

What the four companies said

None of the four argues that the result was impossible. OpenAI accepted the findings and noted that the test was carried out months ago with an earlier version of ChatGPT. It says it has since strengthened its safeguards, including the GPT-6 family, and that it combines internal testing with human review and usage monitoring.

That improvement has not been measured by an independent auditor, and there lies the difference between a response and a test. Anthropic acknowledged that a determined user can bypass safeguards and said it takes the report seriously, but questioned part of the methodology and stated that it could not verify the results on its own. Google noted that older versions of its model were tested.

Fox News reports the same line on Grok, operated by the company the report calls SpaceXAI. Four responses and one shared argument: the April 2026 system is no longer the November one. Comparing the two versions requires trusting the word of whoever makes them.

An audit without recommendations

The report runs to 61 pages and contains no recommendations. That gap says a lot about how GAO works, since it usually audits federal agencies and issues instructions when it finds a failure. Here it examined third-party products and limited itself to documenting a risk so that Congress can decide what to do about it.

Nextgov highlights the unusual nature and adds the methodological caveat signed by the agency itself: the results reflect a one-off assessment, not a representative sample, and anyone repeating the tests today may get different answers. The conclusion GAO does stand behind is more modest and more uncomfortable, because it states in writing that the technology, as it stands on the market today, can be dangerous.

If you work in security, the part that teaches something is the description of the process, not the headline. The platforms did not fail through carelessness: they failed because their defenses activated and could be circumvented with known techniques. It is the same kind of measurement pursued by the automated bug hunting in Linux that OpenAI carried out, with one difference that matters: there the findings are published.

What Hassan wants to legislate

For Hassan, the problem is one of common rules. Companies compete to release models as soon as possible, and her proposal, in the version she has given to the press, is to require that they reach the market with safeguards that cannot be bypassed in this way.

The requirement runs into a practical problem. A safeguard that cannot be bypassed is a promise that is hard to verify, and GAO's test shows that the state of the art consists of getting around them. What can be required is a documented process of adversarial testing before release, with auditable results, and that looks more like what companies already do than like a new law. Every launch comes with its own declaration of safeguards, such as the ones Alibaba presented with Qwen4, and none is measured with the same yardstick.

While Congress decides, pressure comes through another channel: the lawsuit seeking to freeze AI from the large models moves the same debate into the courts, with timelines measured in years. Hassan's committee opens in January 2027, with a new Congress, and hearings serve to stake out positions rather than to write rules.

The problem of proving harm before it happens

The underlying difficulty is not technological but methodological. To prove that a system is dangerous you have to ask it to be, and that request is precisely the one the system is designed to refuse. When the researcher succeeds, they have a proof they cannot publish in full without making the problem worse.

Hence the discomfort of the document. GAO shows that the barrier exists and that it can be crossed, and makes clear that what is new is not the content but the time it takes to turn it into a plan.

For the crypto sector, accustomed to debating whether code is law, the episode leaves a parallel lesson. None of these platforms publishes its weights or its filtering, and yet a regulator measured their behavior with a test lasting hours. The transparency an audit demands is not the transparency a developer demands, and the four companies already know the language in which they will be questioned in January.

José Maldonado

Author

José Maldonado

Activist and blogger on technology, free software and blockchain.

The latest
Anthropic sets the price of its stock market debut at 83 dollars

Keep reading