LLM Jailbreak Research

Large language models are trained to refuse harmful requests and follow the instructions of the people who deploy them. Jailbreak research studies how those guardrails fail. It covers crafted prompts that coax a model past its safety training, as well as prompt injection, where malicious instructions are hidden in documents, web pages, emails, or tool outputs that a model reads while doing its work. The field also includes defensive work: measuring how robust models are, building filters and detection systems, and examining what happens inside a model when it decides to refuse.

Why it matters now

The stakes have changed as language models move from chat windows into autonomous agents. An agent can browse the web, call APIs, and act on a user's behalf. That makes a successful injection more than an embarrassing output. It can lead to leaked data or unauthorized actions. Attacks once demonstrated mostly in research papers are increasingly showing up against deployed systems. Meanwhile, security vendors, standards groups, and AI labs are trying to define what adequate oversight of agentic software should look like. Interpretability methods are also giving researchers new ways to locate where safety behavior lives inside a model, and to see how attacks get around it.

What you'll find here

This hub tracks:

• New attack techniques, including direct jailbreaks and indirect prompt injection
• Defensive research, from mitigation methods to benchmark results
• Agent platform security, including analyses of gaps in commercial agent-building tools
• Interpretability studies that examine the internal mechanisms behind model refusals
• Governance and best practices for overseeing AI agents in production
• Real-world incidents where these attacks have reached live systems