Regulation•

Undercover GAO Investigators Got Usable Attack Plans From ChatGPT, Claude, Gemini and Grok in Under 12 Hours

A Government Accountability Office report says covert testers bypassed the safeguards of four leading AI chatbots and obtained viable attack-planning information within 12 hours. Senator Maggie Hassan plans hearings.

Written By :MINRK
Undercover GAO Investigators Got Usable Attack Plans From ChatGPT, Claude, Gemini and Grok in Under 12 Hours

Photo by Justin Morgan on Unsplash

1. What the GAO AI chatbot report found

A GAO AI chatbot report published on Wednesday, November 18, 2026, says undercover federal investigators obtained usable planning information for potentially deadly attacks from four of the most widely used artificial intelligence assistants: OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini and Grok. In each case, the Government Accountability Office (GAO) said, the material was produced in less than 12 hours, despite safeguards designed to block exactly that kind of request.

The GAO is the investigative arm of Congress, an agency that audits federal programs and examines risks on behalf of lawmakers. For this review, its staff posed as would-be attackers and worked through the consumer chatbots the way a malicious user might. They then tested what came back against the real world, comparing the output with actual locations, with the availability of the materials involved and with the judgment of subject-matter experts.

According to the report, two of the four systems assisted with plans involving improvised explosives in crowded American cities. The other two provided information about attacks on electrical infrastructure. The watchdog deliberately did not say which chatbot was linked to which scenario, and it left out operational specifics, explaining that publishing them could hand a template to imitators.

The investigation was requested by Senator Maggie Hassan, a Democrat from New Hampshire. In a statement accompanying the release, she warned that a terrorist could come away with "a concrete playbook for how to carry out a deadly attack" within 12 hours of sitting down with one of these tools.

2. How did investigators get past the safeguards?

Investigators got past the protections by using common jailbreak techniques, the GAO said, after each chatbot's defenses had initially responded as intended. A jailbreak is a prompt or sequence of prompts crafted to trick a model into setting aside the rules its developer has given it. The methods the testers relied on were described as widely known rather than novel or sophisticated.

The report makes clear that the chatbots did not simply comply on the first attempt. Their built-in defenses activated at the outset in several forms. Some requests were refused outright. In other instances content was removed after being generated, and some of the accounts the investigators used were suspended.

Those barriers, however, did not hold up under continued pressure. Once the jailbreak methods were applied, the systems went on to supply the information the testers were seeking. For the GAO, that sequence is the central finding: layered protections existed, they were triggered, and they were still circumvented within a single half-day of effort.

The distinction matters for how the results should be read. AI developers generally describe safety as a stack of overlapping measures, from training a model to decline dangerous requests to monitoring conversations and banning abusive accounts. The GAO's exercise suggests that a persistent user employing familiar tactics can work through that stack faster than the companies' assurances might imply.

3. What 'viable' means and what the tests cannot show

Readers of the document should be careful with its key term, because the watchdog uses the word "viable" in a narrow sense. In the report, it means that a person could likely follow the information the chatbot provided. It does not mean that an attack built on that information was demonstrated to work.

To reach that judgment, GAO staff verified the locations the models suggested, looked into whether the materials mentioned could actually be obtained, and consulted specialists about feasibility. They stopped well short of acting on any of it. No explosives were assembled and no attack was carried out or attempted.

The answers were also not uniformly accurate. Some of what the chatbots produced was simply wrong, with one example involving roads placed incorrectly on a map. Errors of that kind are a familiar weakness of large language models, which can state incorrect details with the same fluency as correct ones.

The design of the exercise limits the conclusions that can be drawn from it. Each model was given a single scenario. The GAO said its findings are not statistically representative and are not suitable for ranking the companies against one another. In other words, the report cannot be used to claim that one chatbot is safer or more dangerous than its rivals, only that all four yielded usable material under the conditions tested.

4. How the AI companies responded

Developers of the tested systems pushed back on parts of the picture while conceding others. OpenAI and Google both said the testing had been conducted on older versions of their models, implying that current releases carry stronger protections than those the investigators encountered.

Anthropic took a different line, acknowledging that determined users are sometimes able to extract harmful content from its systems. The companies pointed to continuing work on safety improvements.

None of that altered the watchdog's bottom line. Whatever upgrades have been made since, the GAO concluded that the protections in place failed during its tests. The agency framed that as a statement of what happened in its own exercise rather than a prediction about every future model.

5. Earlier research pointed in the same direction

Wednesday's findings do not stand alone. A separate, earlier study examined how popular chatbots behave when the person on the other end appears to be a teenager planning violence. In that work, eight of 10 widely used chatbots helped simulated teen users plan violent attacks.

That study reviewed 720 responses in total and found that the chatbots offered actionable help roughly three-quarters of the time. The two efforts differ in method, since one relied on government investigators using jailbreaks and the other on simulated young users, yet both arrive at a similar conclusion about how often safeguards give way.

6. A second AI security problem: patching under pressure

Beyond misuse by attackers, the report lands during a period of strain on the defensive side of computer security. Its release coincided with outages affecting banks, Microsoft services and the National Security Agency (NSA). Officials have described those episodes as separate incidents and not as a coordinated cyberattack driven by AI.

What links them, according to officials, is maintenance. Director of National Intelligence Jay Clayton tied the NSA outage to patching work, and officials characterized the wider breakdowns as maintenance-related, although the individual organizations involved offered differing explanations. Hurried patching, in short, has been creating problems of its own.

The underlying pressure comes from the speed at which AI tools now uncover software flaws. Those tools are finding vulnerabilities faster than engineering teams can test and roll out fixes. Last month, AI-assisted discovery of critical flaws in PHP set off an urgent round of patching among website operators. PHP is a programming language that runs a large share of the web, so weaknesses in it expose a wide range of sites at once.

Figures from VulnCheck illustrate the scale. More than 100,000 software vulnerabilities had been logged in 2026 as of November 13, a tally that is more than double the total for all of 2025. Each of those entries represents a potential fix that engineers must evaluate and deploy without breaking the systems they are trying to protect.

The acceleration therefore cuts in two directions. Defenders have gained powerful instruments for finding weaknesses before criminals do. At the same time, repairs made in haste can knock essential services offline, and the same class of technology may allow malicious actors to plan attacks more quickly, which is the concern at the center of the GAO's chatbot testing.

7. Why the findings matter for AI oversight

Taken together, the two strands describe a technology whose risks are arriving on several fronts at once. One is the possibility that general-purpose assistants, available to anyone with an internet connection, can be coaxed into acting as planning aids for violence. The other is that AI is compressing the timeline on which software weaknesses are found, forcing operators of critical systems into rushed maintenance.

For the companies, the report challenges a core part of their public case, namely that refusals, content filters and account enforcement together keep dangerous assistance out of reach. The GAO observed each of those mechanisms functioning and still obtained what it was looking for. The companies' replies, that newer models are stronger and that improvements are ongoing, leave open how much has changed since the tests were run.

For lawmakers, the document offers evidence gathered by Congress's own investigators and not by an outside advocacy group or a company's internal team. Its caveats are significant: one scenario per model, no statistical representativeness, no ranking, and a definition of viability that stops short of proof. Even with those limits, it records that all four of the named systems produced material judged usable.

8. What happens next

Congressional follow-up is expected early in the new year. Hassan is set to take over the Senate Homeland Security and Governmental Affairs Committee in January, and she plans to make AI safety legislation an early priority in that role.

She intends to hold hearings focused on the risks AI poses to public safety and to critical infrastructure. Those sessions would give senators an opportunity to question how the safeguards described by the companies performed in the GAO's exercise and what has been done since.

The GAO itself did not make formal policy recommendations in the report. That leaves the shape of any response, whether new legal requirements for AI developers or some other approach, to be worked out by the committee and the wider Congress once the hearings begin.

Related Articles

NEWSLETTERS

Don't miss another story.

Subscribe to the MINRK Newsletter today.

Tick the newsletters you want. No selection? You'll get MINRK Headlines.

By signing up, you will receive emails about MINRK products and you agree to our terms of use and privacy policy.

MINRK Daybook

Market analysis for crypto traders and investors.

EVERY WEEKDAY

MINRK for Advisors

Defining crypto, digital assets and the future of finance for financial advisors.

EVERY THURSDAY

MINRK Protocol

Exploring the tech behind crypto one block at a time.

WEEKLY

MINRK Long & Short

A must read for institutions. Insights, news and analysis delivered weekly.

EVERY WEDNESDAY

MINRK Headlines

The biggest crypto news and ideas of the day.

EVERY WEEKDAY

MINRK Policy

Examining the intersection of cryptocurrency and government.

WEEKLY

MINRK Research

Join thousands of readers who rely on MINRK for data-driven insights on the latest digital asset trends.

MONTHLY