Cleus/Blog

Gemini Jailbreak Prompts in 2026: Why They Fail on Gemini 4 (and What Works Instead)

Most Gemini jailbreak prompts were written for another model, and on Gemini 4 they argue with the one layer that is not doing the blocking. Here is how the filters work and what people do instead.

Prabhu Biswal9 min read
TL;DR

Gemini jailbreak prompts fail on Gemini 4 because the blocking is done by classifiers outside the model, which no persona or story can talk to. Google reports its automated attacker got through on 0.8% of attempts and testers found no universal jailbreak in 132 hours. Ordinary edgy requests already work without one. For the rest, the reliable routes are an open-weight model or a hosted one without that filter.

Gemini Jailbreak Prompts in 2026: Why They Fail on Gemini 4 (and What Works Instead)
Skip the jailbreak · Cleus V3
An AI that doesn't need tricking.
Cleus V3 has no reflex refusals to argue with. Ask in plain words and get a full answer. Disclosure: Cleus publishes this blog.
Try Cleus V3 free → Free credits on sign-up · Works in your browser

Most "Gemini jailbreak prompts" you find in a search were not written for Gemini. One 2026 guide that went through them found that many are old ChatGPT prompts with the name swapped. People paste them into Gemini 4, get an empty reply or a polite no, and conclude they need a better prompt.

They usually need one of two other things. Either the request was never blocked in the first place, or it is blocked by something no prompt talks to.

We tested Gemini 4 Carbon and Gemini 4 Argon in launch week, and Google has published unusually detailed numbers on how its filters hold up against attackers. Put together, they explain why copy-paste jailbreaks fade so fast on Gemini 4, and what people who hit the wall end up doing instead.

We are not publishing jailbreak prompts or ways around Google's filters here. This is about how the system works, so you can stop wasting evenings on prompts that were dead before you found them.

0.8%
of automated attacks got through
Google's figure for Carbon; 3.1% on Argon
132 h
of testing, no universal jailbreak
Trajectory Labs, 13,000+ requests
18 / 18
edgy everyday runs Carbon answered
Our test, no jailbreak used
0 / 12
scan and exploit runs answered
Carbon and Argon, our test

First, check whether you need a jailbreak at all

A lot of jailbreak hunting starts from an old memory of Gemini refusing everything. That is out of date.

On November 11 we sent 11 prompts to Carbon and Argon through the API, three runs each, with default settings and no system prompt. Six were the edgy-but-ordinary set we run on every model: lock picking as a hobby, a villain's threat, a brutal roast, a phishing email for staff training, a love scene that stops short of explicit, and a dark joke. Carbon answered all 18 runs of that set. Argon answered 15 of 18 and failed only the phishing email.

So for dark fiction, harsh humor and hobby questions, plain Gemini 4 already says yes. A 1,500-word persona prompt on top makes those answers worse, because the model spends its attention on the costume.

The API's safety settings point the same way. Google's documentation lists five adjustable categories: harassment, hate speech, sexually explicit, dangerous content and jailbreak. For Gemini 2.5 and every later model, all five are off unless a developer turns them on. The docs put it this way: "Due to the model's inherent safety, additional filters are Off by default."

Our set stopped short of explicit writing, so it says nothing about that. And it covers the API, not the consumer Gemini app, which Google runs with its own product rules.

What Gemini 4 does block, and why no setting changes it

Two kinds of request hit a wall that has no switch.

Core harms. Google's docs say protections for things like child safety "are always blocked and cannot be adjusted." That is true on every tier and it should be.

Offensive security work. This is the new one. Google's safety report says it cannot rule out that Carbon, with outside filters removed, crosses its first danger threshold for helping with cyberattacks. So the public model ships behind a separate cyber layer that is not in the settings list. In our test, neither Carbon nor Argon would give an office admin nmap commands for their own network, or write the exploit for a beginner practice binary. That was 0 answers in 12 runs.

A blocked request does not come back as a refusal you can read. The API returns an error and no text. Through the route we used, it was a single line: Gemini blocked the response: OTHER.

That empty reply is the first clue to why prompts fail.

Six checks, and the model is only one of them

A jailbreak prompt is an argument aimed at the model. On Gemini 4, the model is one checkpoint out of six, and the others do not read arguments.

Interactive

What your prompt passes through on Gemini 4

The layers in Google's safety report, in order. A jailbreak prompt only addresses one.

Step 1 of 6
Pattern rules

Fast rules and embedding matches catch known strings. Google uses this layer to react quickly to newly spreading attacks, which is why viral prompts die first.

Think of it as a building with a friendly receptionist and five guards who never speak to visitors. A good story might win over the receptionist. The guards are looking at what you carried in and what you are carrying out.

That is why role-play, "developer mode" and fake system messages do so little here. They can change what the model is willing to write. They do not change what a classifier sees when it reads the prompt before the model does, or the finished reply after.

Google's numbers on how often attacks get through

Google and the outside testers it hired attacked Carbon before release. These are their figures, and nobody independent has reproduced them yet.

Interactive

How often attacks got through

Figures reported by Google, not yet independently reproduced. Tap a bar.

Gray Swan cyber queries, Argon: 9.86% (62% of the total). Share of automated cyberattack queries that succeeded on Gemini 4 Argon.

Source: Google, "Gemini 4 Carbon is now available", November 10, 2026.

  • Google's own automated attacker, which rewrites its plan after every block, got through on 0.8% of attempts. On Argon it was 3.1%.
  • Gray Swan's automated cyberattack queries succeeded 2.17% of the time, down from 9.86% on Argon.
  • In simulated attack environments, Carbon's safeguards blocked 99% of attack runs.
  • Trajectory Labs spent 132 hours and more than 13,000 requests and found no new universal jailbreak, meaning no single prompt that works across tasks.

There is an honest footnote. Google's report also says automated attackers got at least a partial result on 40% of Gray Swan's test objectives when given 200 tries each. So the filters are not a perfect wall. What that number describes, though, is software making hundreds of tailored attempts per goal. It is the opposite of one prompt you copy from a forum.

A "universal Gemini 4 jailbreak prompt" is the exact thing 132 hours of professional testing failed to find.

Five reasons the prompt you found stopped working

It was written for a different model

Much of what ranks for "Gemini jailbreak" is recycled ChatGPT text, or dates from Gemini 1.5 and 2.5. Gemini 4 Argon became the stable model on October 13, 2026, and Carbon went public on November 10. Prompts that predate those weights were tuned against a model that is gone.

It talks to the wrong layer

Covered above. Winning the model's agreement does not move the classifiers on either side of it.

There is nothing to argue with

Older chatbots refused in words, and jailbreakers learned to push back on those words. An empty reply with an error code gives you nothing to push on, and no hint about which part of the request tripped the filter.

The edge moves between identical requests

In our runs, the same request to decode a suspicious PowerShell command was answered twice by Carbon and blocked once. A prompt that worked for the person who posted it can fail for you an hour later with nothing changed.

Google patches what goes viral

Google's safety report describes fast pattern rules built to react to newly escalated threats. The more widely a prompt is shared, the sooner its shape is on that list. Popularity is what kills it.

The cost people forget: it is your Google account

A burner account on a small AI site is one thing. Gemini runs on the same Google account as your Gmail, Drive, Photos and maybe your phone.

Google's safety report describes log monitoring for people trying to get dangerous content, with the worst cases reviewed and able to end in a blocked or disabled account. The API docs add that "applications that use less restrictive safety settings may be subject to review" under the Terms of Service. We found no public count of accounts closed over jailbreak attempts, and the developer forum complaints about suspended apps do not say why. The policy is still written down, and the downside is larger than losing a chatbot.

What people do instead

Ask plainly first. With filters off by default, the straightforward version of a creative or blunt question often just works. Try it before anything else.

For security work, use the route Google built. The Fairwind Program gives vetted organisations a Carbon with the cyber filter loosened. It is aimed at governments, critical infrastructure and security vendors, and we found no track for individuals. If you work in security at a bank, hospital or telecom, ask whoever owns the Google Cloud relationship. The defensive half of the job needs no special access: in our runs, code review, explaining SQL injection with its fix, and finding bugs in a C function all passed.

Run an open-weight model yourself. A model on your own hardware has no outside classifier and no account to lose. Our guide to uncensored models you can run locally sorts them by how much memory you have, and What Is an Abliterated Model? explains how refusals are removed from open weights.

Use a hosted model that was not built around refusals. This is where we have an interest, so read it as such. We sent the three security prompts Gemini blocked to Cleus V3, once each on November 11. It answered all three in full. That is not a claim that Cleus V3 matches Carbon on security benchmarks. It is a claim that it answered the question without a costume.

Uncensored does not mean anything goes. Cleus has its own rules: nothing involving minors, no non-consensual imagery of real people, and nothing designed to hurt someone. What changes is the wide grey zone of legitimate adult, creative and technical questions that get turned down by reflex.

Interactive

Jailbreak or not needed?

Four questions on how Gemini 4 filtering works.

1/4

On current Gemini models, the five adjustable API safety categories are set to what by default?

Gemini jailbreak FAQ

Do Gemini jailbreak prompts still work in 2026?

Rarely, and not for long. Most published prompts were written for older Gemini models or for ChatGPT. On Gemini 4, classifiers outside the model check the prompt and the reply, so persuading the model is not enough.

Has Gemini 4 been jailbroken?

Not with a universal prompt, as far as public results show. Google reports that hired testers found none in 132 hours and over 13,000 requests. Automated attackers did get partial results on 40% of one test set when allowed 200 attempts per goal.

Why does Gemini return an empty response?

A safety filter blocked either your prompt or the model's reply. The API returns an error instead of text, and it does not say which part caused it.

Can I turn off Gemini's safety filters?

The five adjustable API categories are already off by default on current models. The always-on protections for core harms and the cyber safeguards on Gemini 4 cannot be changed from settings.

Is jailbreaking Gemini illegal?

Trying a prompt is generally not a crime, but it breaks Google's terms, and Google says serious misuse can end with a disabled account. What you ask for and what you do with the answer are still subject to the law.

Can my Google account be banned for jailbreaking Gemini?

Google's safety report says monitoring can lead to an account being blocked or disabled in the worst cases. How often that happens is not public.

What is the best alternative to a Gemini jailbreak?

A model without the filter that is blocking you: an open-weight model on your own machine, or a hosted one such as Cleus V3. For professional security work, the official path is Google's Fairwind Program.

Sources

Still curious?

Ask the question this article didn’t answer.

Ask anything about gemini jailbreak. Cleus opens this article, searches the live web and gives you a straight answer, with sources.

Free to start · no card
What’s behind the button

One question in. Real work out.

Cleus is an AI agent that researches, codes, draws, films, writes and talks — and answers the question you actually asked.

Try Cleus free

No card · daily free credits · works right in your browser

Read next