Top news
Investing in the public benefit OfficialAnthropic NewsOct 18Popular Musk Pledges SpaceX Shares to ‘Every Kid in America’ WebBloomberg TechnologyOct 19Popular Websites are closing their doors to AI agents WebThe Verge AIOct 17 This is the new ChatGPT Voice, powered by GPT Live OfficialOpenAI on YouTubeOct 18 How AI giants will reshape global philanthropy WebAxios AIOct 19 Gemini 4 Argon Really Surprised Me WebTwo Minute PapersOct 19 Google Gemini 4 agents just beat Minecraft with no human intervention WebEngadget AIOct 19 Who could it be? 🧐 OfficialGoogle on YouTubeOct 18 Anthropic gives outside AI watchdog “employee-like” access WebArs Technica AIOct 18 What are self-replicating ‘AI worms’, and why is Anthropic worried about them? WebThe Guardian AIOct 18
GlossarySafetyIntermediate

Jailbreaks

Also written: Prompt Jailbreaking, Universal jailbreaks, AI jailbreaks, Jailbreakability, Universal jailbreak attacks, Jailbreak demonstration, Jailbreak-like instructions

A security exploit that bypasses AI model restrictions to make it perform unintended or prohibited actions.

In plain terms

Jailbreaks are methods people use to trick AI models into breaking their own rules. Think of it like finding a loophole in the AI's programming that lets it say or do things the developers didn't intend.

For example, a user might try to convince an AI assistant to generate harmful content by framing the request as a fictional story or a hypothetical scenario. The AI, trying to be helpful or follow the user's instructions, might then bypass its safety filters.

The expert version

Jailbreaks are adversarial attacks designed to circumvent the safety guardrails and content policies implemented in large language models. These techniques exploit vulnerabilities in the model's training or fine-tuning to elicit responses that are harmful, unethical, or otherwise prohibited by the developers. Common methods include prompt injection, role-playing scenarios, and exploiting the model's tendency to follow complex instructions.

Why it matters now

Frontier AI labs are increasingly disclosing safety incidents and developing new systems to track and report unexpected AI behavior. This focus on model misbehavior and the ongoing efforts to test and verify AI safety indicate a continued challenge in preventing AI models from being tricked into violating their intended use or safety rules.

Where it shows up 26 items, newest first
How each company uses it
Anthropic Claude
Anthropic is developing an AI audit service to help enterprises verify the safety and reliability of AI agents, and its models are tested for resistance to misuse and jailbreaking.
Google Gemini
Google Gemini's development of advanced AI for complex tasks includes enhancing cybersecurity with frontier-level defenses, implying a need to protect against exploits.
OpenAI ChatGPT
OpenAI is implementing faster incident reporting and new disclosure systems to track and reveal instances of AI models misbehaving and finding new ways to break rules.
xAI Grok
xAI Grok is mentioned in the context of new model releases and competition, suggesting its models are also subject to the same safety and security considerations as other frontier models.
Related terms

Prompt InjectionAdversarial AttacksAI SafetyRed TeamingContent ModerationGuardrails

Often confused with
Prompt Engineering
Prompt engineering focuses on crafting effective prompts to
Fine-tuning
Fine-tuning is a training process to adapt a model for specific tas
Model Hallucination
Hallucinations are instances where an AI generates factuall
Origin

The term 'jailbreak' originates from the practice of bypassing security restrictions on devices like smartphones to gain unauthorized access or functionality. In AI, it was adopted by the security and AI research communities to describe similar bypasses of model restrictions.

Try this about 5 minutes

To see a simple form of this, try asking an AI assistant to write a story where a character performs a forbidden action. Then, try to rephrase the request to make it sound like a hypothetical or a fictional scenario. Observe if the AI's willingness to generate the content changes based on how the prompt is framed.

Try it: Playbook entries that use thisPlaybook →

The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.

Questions people ask

Are jailbreaks illegal?
Jailbreaking itself is not illegal, but using it to generate illegal, harmful, or unethical content can have consequences depending on the platform's terms of service and the nature of the content.
How do AI companies prevent jailbreaks?
Companies use various methods, including extensive safety training, content filters, prompt analysis, and red teaming to identify and patch vulnerabilities that allow jailbreaks.

First seen on Subvolts Apr 19, 2024. Updated Oct 13, 2026. Written by an AI model from the items above and reviewed against them weekly; report an error.