GAO jailbreaks ChatGPT, Claude, Gemini and Grok into terror plans in under 12 hours

In spring 2026, investigators from the US Government Accountability Office (GAO) got past the safeguards of four major AI chatbots and came away with workable plans for bombings and attacks on the power grid. The companies say older model versions were tested. Senator Maggie Hassan wants a law that sets stricter safety requirements.

Industry Nov 18, 2026 6 min read Source: Brave News
Illustrative image, AI-generated

What's going on

On Wednesday, November 18, 2026, the Government Accountability Office (GAO), the nonpartisan auditing arm of the US Congress, released a report that makes uncomfortable reading for the big AI companies. In spring 2026, undercover GAO investigators tested the four best-known AI chatbots: OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini and SpaceXAI's Grok. All four could be pushed into producing detailed terror attack plans that experts judged workable. Each one took less than 12 hours.

The investigation was requested by Democratic Senator Maggie Hassan of New Hampshire. Fox News, ABC News, NBC News and the New York Post are among the outlets that have reported on it.

How the test worked

According to ABC News, in April the investigators randomly assigned each platform one of two scenarios: an attack with an improvised explosive device (IED, meaning a homemade bomb) or an attack on the US power grid. Each platform was tested exactly once. The GAO deliberately did not disclose which platform got which scenario or which one produced which result. Specific findings and model details were redacted from the report so nobody can repeat the experiments.

At first, all four chatbots triggered their safeguards and blocked or removed harmful content. But the investigators still got through using what the report calls "common techniques to circumvent safeguards." These are jailbreaks: prompts worded so that the model ignores its built-in rules. A jailbreak might be a made-up role-play scenario, or a task broken into many harmless-looking steps. The GAO doesn't say which techniques it used. NBC News describes it as a "jailbreak test" that showed how easy the safety measures are to defeat.

What the chatbots produced

According to ABC News, two platforms each drew up their own plans for building, placing and setting off bombs in major cities. They suggested specific dates, times and targets around scheduled public events, including a hotel parking garage and a convention center. Both also gave tips on how to avoid drawing police attention. One platform even laid out a timeline for preparing and carrying out the attack, including a shell company and purchases spread across several states. The GAO writes:

"At least one platform provided enough information that could help a malicious actor evade detection and build a very large explosive device."

The other two platforms described attacks on electrical substations in major US cities, aimed mainly at causing economic damage. The chatbots named specific drone and weapon models and explained how to get hold of them. ABC reports that they also generated a map and an image showing how the attack would play out.

The report says each platform "provided detailed information that could be used for attacks, such as specific targets, locations, timing and methods of attack."

How the GAO checked the plans

The agency didn't take the answers at face value. It checked whether they would actually work. According to ABC News, it did this through site visits, its own research and interviews with experts, including the head of a fire department bomb squad in a "large metropolitan area." The New York Post reports that law enforcement agencies were consulted too. The GAO concluded that the information from all four platforms was feasible enough to help someone plan an attack.

Kevin Walsh, the GAO's director of IT and cybersecurity audits, told reporters that the companies had assured the agency they take the issue seriously and are continuing to research it.

"However, we conclude that this technology, in its current state, can be dangerous, and its misuse poses a growing risk to our country."

The report doesn't say whether the tests led any of the companies to alert police. Some platforms did suspend the investigators' accounts after the study.

How the companies responded

All four companies told the GAO that they test their systems continuously, use several layers of safeguards and have improved their models since the spring. Their responses differ in the details:

Company Platform Response (as reported)
Anthropic Claude Acknowledges that determined users can get around safeguards and says it takes the findings seriously. But it questions parts of the methodology and says it can't independently verify the results.
Google Gemini Points out that older model versions were tested: "We continue to improve our models."
OpenAI ChatGPT Calls the findings concerning but stresses that the test ran on an older version of ChatGPT. It says there have been significant improvements since then, including with the GPT-6 model family, and that it relies on monitoring, human review and enforcement.
SpaceXAI Grok Says it is working "diligently" to improve safety with every new model version.

Because the GAO won't say which platform produced which result, you can't match these statements to specific test outcomes.

The political backdrop

The report arrives as Washington pays closer attention to AI misuse. ABC News notes that, according to authorities, attackers used AI to plan shootings in Florida and Canada. The New York Post cites a New York Times report that the terror group Boko Haram has used AI tools in combat. In September, Anthropic itself disclosed that Claude had been misused in China, Russia and Yemen to build software for developing firearms, missiles, armed drones and bombs. ABC reports that users in Yemen failed in their attempts to build missiles.

Maggie Hassan becomes chair of the Senate Committee on Homeland Security and Governmental Affairs in January, when Democrats take the Senate majority. She has announced hearings on AI safety for 2027, including on risks to critical infrastructure, and plans to push a bill with stricter safety requirements.

"The threat AI poses to public safety is no longer theoretical. If Congress doesn't act, it's a question of when, not if, an attacker uses AI to plan and carry out a major attack."

Hassan says the companies face "extraordinary competitive pressure" to ship their models as fast as possible. To level the playing field, she argues, models should be required to have safeguards that can't be bypassed this way before they're released.

Our take

It's not news that chatbots can be tricked with jailbreaks. Security researchers have been showing that for years. What's new is that a government auditor tested all four leading platforms at once, in a systematic way, and that experts judged the results workable in practice. That gives the US debate over mandatory pre-release safety testing real political weight.

Some questions remain open, though. Each platform was tested only once, using spring model versions that have partly been replaced. The methodology is largely redacted. That makes Anthropic's objection that it can't verify the results understandable, even if the redactions are meant to stop copycats. It's also unclear how much the AI actually added beyond what's freely available elsewhere. That question of "uplift" for attackers is central in security research. The companies' claims about newer models like GPT-6 come from the companies themselves, and nobody has checked them independently. One thing is clear: even with several layers of safeguards, determined attackers can apparently get what they want in hours, not weeks.

Sources

#ai safety#jailbreak#regulation#usa#chatbots
Read the original source

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.