US government auditors got ChatGPT, Claude, Gemini and Grok to produce attack plans: what it means
The GAO, the watchdog that audits the US government, posed as ordinary users and got the four most popular chatbots to help plan an attack in under 12 hours. Here's what they did, how the companies responded and what it means for you.

The Government Accountability Office (GAO), Congress’s nonpartisan watchdog, has published an awkward report for the AI industry. Its investigators posed as ordinary users and got ChatGPT, Claude, Gemini and Grok to hand over information that could help plan an attack. Each one took less than 12 hours.
The report was made public on November 18, 2026. Here’s what the auditors did, what the findings don’t mean and how the companies answered. We won’t share any details on how it was done, and neither does the report.
At a glance
| Who | GAO, the US Congress’s audit office |
| Requested by | Sen. Maggie Hassan (D-NH) |
| When tested | Spring 2026 (report dated Oct 16, released Nov 18) |
| Which chatbots | ChatGPT, Claude, Gemini and Grok, the four most used as of April 2026 |
| Result | All four eventually produced information useful for planning an attack |
| Time | Under 12 hours each |
| Recommendations? | None: the report only describes what happened |
What the investigators did
- They set up anonymous accounts, like anyone would, starting on free plans and paying for upgrades when they hit message limits.
- With expert help, they picked two attack scenarios: one involving an explosive device and one targeting the power grid. Each chatbot was randomly given one.
- At first, all four pushed back: they refused, removed answers or even suspended accounts.
- The investigators then used well-known tricks to get around those safeguards (so-called jailbreaks) and, with persistence, got what they were after.
- Finally, they checked with experts, site visits and research whether the plans could actually be carried out. No attack was attempted, of course.
The report doesn’t say which chatbot did what: it calls them Platforms A, B, C and D. So it can’t tell you which one is “safest.”
The important caveats
To read it fairly, it helps to know what the report does not show:
- The information wasn’t new. GAO itself says nearly all of it was already online. What the AI did was gather and organize it in hours instead of weeks or months.
- The plans had mistakes. Techlicious reports misplaced roads on a map and a suggested venue that had no event on the proposed date. AI makes things up here too.
- It’s a tiny sample. One test at one point in time, with no control group. GAO says so.
- The models have changed. Testing took place in spring, and these models are updated every few weeks.
What the companies say
All four companies sent written comments, which GAO published with the report. Broadly, they all say they keep improving safety.
- OpenAI (ChatGPT) says GAO tested an earlier version and that its current model refuses the requests it judged fell short of its rules, according to Techlicious.
- Google (Gemini) also points out that the April tests used older models.
- Anthropic (Claude) says its defenses assume any single safeguard can be broken, and that it has expanded them since. Per Decrypt, it acknowledged that determined users can sometimes extract harmful material.
- SpaceXAI, which the report names as Grok’s developer, says it has improved every layer of its safety systems and asked GAO for the chat logs.
What happens next
Sen. Hassan, who requested the audit, is set to chair the Senate Homeland Security and Governmental Affairs Committee from January 2027. She says the danger is “no longer theoretical,” per Techlicious, and plans an early hearing on AI risks. Her priorities include more rigorous safety testing before new models are released.
In Europe, the EU AI Act already requires the most powerful models to assess and reduce risks like these.
What it means for you
In everyday use, nothing changes. But the report carries three useful lessons:
- Chatbot filters aren’t a wall. They stop most people, not someone with time and bad intentions.
- An AI saying it doesn’t make it true. Even the “viable” plans had errors.
- If you see someone using AI to plan harm, report it to local law enforcement (in the US, the FBI takes tips at tips.fbi.gov). Most chatbots also have a button to report conversations.
Note: we read GAO’s official report page and part of the full report. The company responses and the senator’s comments are cited from Decrypt and Techlicious.
Sources
- GAO: Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans (GAO-27-108814, released Nov 18, 2026)
- Decrypt: ChatGPT, Claude, Gemini and Grok Helped Plan Terror Attacks in Government Tests (Nov 18, 2026)
- Techlicious: Federal test got attack plans from ChatGPT, Claude, Gemini and Grok (Nov 19, 2026)


