GAO covert tests got ChatGPT, Claude, Gemini and Grok to produce viable attack plans
A GAO report released on 18 November 2026 says covert testers bypassed safeguards on ChatGPT, Claude, Gemini and Grok and got viable attack plans from all four.
By Emre A.Published
Tech AI LatestThe US Government Accountability Office released a report on 18 November 2026 saying its covert testers got ChatGPT, Claude, Gemini and Grok to produce viable plans for attacks in the United States. The chatbots first refused, but common jailbreak techniques got past the safeguards. The tests ran in spring 2026 on models that have since been replaced. Senator Maggie Hassan, who requested the work, plans a January hearing on AI safety.
Key facts
- Report: GAO-27-108814, released 18 November 2026 (GAO report)
- Platforms tested: ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Grok (xAI, now SpaceXSI)
- Result: all four produced attack-planning information after initial refusals were bypassed; two plans involved a mass-casualty bombing, two targeted the electrical grid
- Timing: tests ran in spring 2026; outlets report each platform produced a plan in under 12 hours
- Requested by: Senator Maggie Hassan, expected to chair the Senate Homeland Security Committee from January
- Not stated: which platform produced which plan, and the specific jailbreak techniques
What the GAO found
The Government Accountability Office, the investigative arm of the US Congress, ran covert tests in spring 2026 to see whether widely available chatbots could help someone plan an attack. Its report, GAO-27-108814, titled "Covert Testing Used Common Platforms to Create Viable Attack Plans", was released on 18 November. The testers used ChatGPT from OpenAI, Claude from Anthropic, Gemini from Google and Grok from xAI.
According to the Washington Post, the Washington Times and other outlets that reviewed the report, each chatbot at first refused the requests. Investigators then used common jailbreak techniques, which the report does not name, and all four platforms produced attack-planning information. Two platforms produced plans for a bombing intended to cause mass casualties and two produced plans against electrical substations, with targets, locations and methods. Tech Times and WMUR report that each platform reached a plan in under 12 hours. The report does not say which chatbot produced which plan, and outlets report that the GAO says its results cannot be used to rank the platforms' safety.
The GAO also notes limits. Some answers contained errors, the investigators did not build anything or carry out an attack, and the underlying information was not new. Its conclusion, as quoted by several outlets: even with safeguards in place, there are no current generative AI systems that are completely immune to such misuse.
How the companies responded
All four developers said the GAO tested models that have since been replaced. Anthropic said it takes the findings seriously and does not dispute that a determined tester can extract harmful content by misrepresenting a request's purpose, but said it could not verify the results, that its safeguards assume any single protection can be defeated, and that the tests reflect its system as of May. OpenAI said its newer GPT-6 Astra refused the requests tied to the three ChatGPT responses it found fell short of its policies, and objected that the report does not always separate ChatGPT's output from the investigators' own sources. The GAO replied that consulting outside sources matches how a malicious actor would work. Google said its work on safety and security is never complete; xAI said it improves safety with each release.
Why it matters and what comes next
The report puts a congressional watchdog's own test results next to the labs' self-assessments, and it lands as the labs rate their own models more dangerous. OpenAI's GPT-6.2 system card places both new models at its Critical cybersecurity level, and Anthropic sells the less restricted Claude Mythos 5.2 only to vetted organisations. Anthropic's own threat report in September described real misuse it had disrupted. The GAO did not test those newer models, so it cannot say whether the newer safeguards would have held.
Senator Hassan, who requested the work and is expected to chair the Senate Homeland Security and Governmental Affairs Committee in the new Congress, has scheduled a hearing on AI safety for January and has said AI safety legislation will be an early focus. Current US rules mostly require reports after harm occurs, as this site's comparison of incident reporting laws shows. More coverage is on the cybersecurity hub.
What changed
| Before | Chatbot safeguards against attack planning were tested mainly by the developers and contracted red teams |
|---|---|
| After | A GAO audit reports that all four major chatbots produced viable attack plans in covert tests in spring 2026 |
| Who it affects | OpenAI, Anthropic, Google and xAI; Congress and federal agencies; emergency planners |
Questions people ask
- Which chatbots did the GAO test?
- ChatGPT, Claude, Gemini and Grok, in covert tests in spring 2026. All four produced attack-planning information after their first refusals were bypassed, according to report GAO-27-108814.
- Did the GAO say which chatbot was the most dangerous?
- No. The report does not match plans to platforms and says its results cannot be used to rank the platforms' safety.
- What did OpenAI and Anthropic say?
- Both said the tests used older models. OpenAI said GPT-6 Astra refused the relevant requests; Anthropic said it could not verify the findings and has strengthened its safeguards since May.
Sources
- GAO-27-108814, Artificial Intelligence: Covert Testing Used Common Platforms to Create Viable Attack Plans (US Government Accountability Office (via Senator Maggie Hassan), 18 Nov 2026)
- Chatbots helped plan terrorist attacks in tests by congressional watchdog (The Washington Post, 18 Nov 2026)
- Watchdog says it tricked top AI chatbots into building bomb and grid-attack plans (The Washington Times, 18 Nov 2026)
- Government watchdogs asked chatbots to plan a terrorist attack. They did (WMUR, 18 Nov 2026)
- GAO Bypassed AI Chatbot Safeguards on ChatGPT, Claude, Gemini, Grok; Attack Plans in Under 12 Hours (Tech Times, 19 Nov 2026)
Reference pages
- OpenAI company
- Anthropic company
- Google company
- xAI (SpaceXSI) company
Spotted an error? See the corrections policy and tell the editor. Facts in this story come from the sources listed above; vendor claims are attributed.
Comments
Loading comments.





