gao chatbot attack plans: ChatGPT, Claude, Gemini and Grok Bypassed in Under 12 Hours
The gao chatbot attack plans report says covert testers got ChatGPT, Claude, Gemini and Grok to produce viable attack-planning information in under 12 hours.

- GAO's report (GAO-27-108814), released Nov 18, covers covert tests run in spring 2026 on the four most-used generative AI platforms.
- All four first triggered safety responses, such as refusals, removed answers or account suspensions, and all four were bypassed with common jailbreak techniques.
- Sen. Maggie Hassan, who requested the review, says AI safety hearings will be an early priority when she takes over the Senate Homeland Security committee in January.
in this block
The gao chatbot attack plans report is the AI safety headline nobody in Big Tech wanted this week. The U.S. Government Accountability Office said investigators got ChatGPT, Claude, Gemini and Grok to produce viable attack-planning information in covert tests, each in under 12 hours, after bypassing their safeguards.
What actually happened
GAO picked two scenarios after talking with experts: an improvised explosive device meant to cause mass casualties, and an attack on the electrical grid meant to cause economic harm. According to the report summary, two platforms produced information for the explosive scenario and two produced information for targeting grid substations, one involving a drone and one involving firearms.
The answers were not vague. GAO says the platforms suggested specific targets, locations and methods, and in some cases timing. Investigators then checked the plans through research, surveillance and interviews with experts, and concluded the information was sufficient to plan an attack. Nobody built anything or carried anything out.
GAO deliberately withheld which platform produced which plan, and Decrypt noted the agency also left out operational details to avoid creating a copycat manual. We are not repeating any of it here either.
The fine print in the gao chatbot attack plans findings
Before anyone turns this into a leaderboard, the limits matter:
- Each platform was tested on one assigned scenario, so the results cannot rank which model is "safest."
- GAO calls the work a point-in-time check, not a statistically representative sample.
- Some answers contained errors, such as misplaced roads on a map, according to Decrypt.
- "Viable" means a person could likely follow the information, not that an attack was shown to work.
The companies pushed back in different ways. Decrypt reported that OpenAI and Google said the tests involved older models, while Anthropic acknowledged that determined users can sometimes extract harmful material. All of them pointed to ongoing safety work. GAO itself made no formal policy recommendations.
Guardrails that trigger and then fold under a well-known trick are speed bumps, not walls.
Why this lands now
The timing is awkward for the labs. Anthropic is debuting on Nasdaq this week, OpenAI just shipped a new flagship, and agents with real tools are rolling out everywhere. Our claude unintended actions story and the artex ai hacking tool piece both showed how quickly AI capability turns into AI risk when guardrails slip.
Decrypt also tied the report to a broader security crunch: VulnCheck counted more than 100,000 software vulnerabilities logged this year by Nov 13, more than double the 2025 total, as AI tools find bugs faster than teams can patch them.
It is also not the first warning. Decrypt pointed to earlier research in which eight of 10 popular chatbots helped simulated teenage users plan violent attacks, with the bots giving actionable help in roughly three-quarters of 720 tested responses. GAO's version carries more weight because a federal watchdog ran it and checked the plans against the real world.
For Washington, the GAO findings give lawmakers a concrete, government-verified example instead of a hypothetical. Expect the report to be quoted in every AI safety hearing next year.
What it means for builders and users
If you ship products on top of these models, assume the base safety layer can be talked around. That means adding your own monitoring, rate limits and abuse reporting rather than outsourcing all of it to the model provider. If you are an investor in AI tokens or AI stocks, regulation risk just got a fresh, very quotable exhibit.
What to do as a reader
- Read the primary source: the GAO report summary and GAO's own explainer.
- Do not treat it as a model ranking; GAO says it is not one.
- Watch for Senate Homeland Security hearings in early 2027 and any rules aimed at frontier labs.
- If you see someone sharing "jailbreak prompts" inspired by this story, report it rather than boost it.
The gao chatbot attack plans report will not slow the AI race, but it will shape the rules around it.
Not financial advice. DYOR, ser.