I Jailbroke Gemini 3.5... And It Actually Worked

Added: 2026-06-04

[00:00:00][music] >> Hello everyone and welcome back to the channel. First of all, thank you all for the incredible support and love you shown on my previous videos. I truly appreciate every like, comment, and subscription. It motivates me to keep creating more cybersecurity and security research content. In this video, I'll be demonstrating a jailbreak that I discovered in Gemini 3.5 flashlight. The purpose of this demonstration is to show how the model can be manipulated into bypassing its intended safety restrictions under certain conditions.

[00:00:35]Before we begin, this video is intended for educational and security research purposes only. As you can see here, the flashlight model is selected.

[00:00:45]>> [music] >> Now, let's try a few prompts without any jailbreak.

[00:00:49]>> [music] [music] >> As you can see, the model responds [music] according to its safety policies and refuses requests that violate its restrictions. This confirms that the safeguards are active and that we're starting from a normal, unmodified state. Once we've established this baseline behavior, we'll move on to the jailbreak prompt and observe whether the model's responses change after the prompt is introduced. This step is important because it allows us to clearly compare the model's behavior before and after the jailbreak attempt and determine whether a genuine bypass has occurred.

[00:01:32]>> [music] [music] [music] >> Now, let's move on to the jailbreak itself. First, delete the current chat and start [music] a completely new conversation. This ensures that no previous context affects the results and that we're testing from a clean state.

[00:02:01]Next, open the jailbreak prompt. I've included a link to it in the video [music] description below. You can also see the prompt on the screen right now.

[00:02:09]Just a quick note, I am not the original author of this jailbreak prompt. I'm demonstrating and testing it as part of security research. Credit goes to the original creator. Once you have the prompt copied, head over to Gemini and make sure that the Gemini 3.5 flashlight model is selected. This is important because the results shown in this video were observed specifically [music] on flashlight. After confirming the model selection, simply paste the jailbreak prompt into the chat [music] box and submit it. Now let's see how the model responds. As you can see, the jailbreak appears to be working. The model has accepted the prompt and its behavior is now noticeably different from what we observed earlier in the baseline tests.

[00:02:50]Now let's try a few test prompts [music] to confirm.

[00:02:57]>> [music] >> As you can see, the model is now generating the code instead of refusing the request. [music] Let's continue testing and see what other prompts the model [music] responds to.

[00:03:24]>> [music] [music] [music] >> Let's ask it what the generated code is actually capable of doing.

[00:03:45]>> [music] >> And one more thing, I'm seriously considering starting a dedicated series on medical [music] device penetration testing and security research. We'll look at how medical devices work, common [music] security issues, attack surfaces, real-world vulnerabilities, responsible disclosure, and the security challenges [music] facing modern healthcare technology. If that's something you'd like to see on this channel, please leave a comment and share your thoughts. Your feedback will help me decide what content to create next.

[00:04:15]>> [music] [music] [music] [music] [music] [music] [music] [music] [music] [music] [music] >> Mhm.