GetJailbroken: A Gamified Playground for AI Agent Security

A new web-based tool called GetJailbroken offers developers a hands-on way to probe the security boundaries of AI agents. The site presents a series of escalating challenges (Levels A, B, and C) where users interact with an AI model and its associated agents, attempting to elicit unintended behaviors or bypass safety measures.

Designed as a practice playground for AI security, the interface features a chat-style interaction grid where users can type messages and observe the agent's responses. A "my grid" display shows the current state, and users can send messages, copy results, or continue experimenting. The goal is to "jailbreak" the agent—meaning to get it to perform actions or generate outputs that violate its intended restrictions.

While the tool does not specify which underlying model or agents it uses, it reflects a growing trend: as developers increasingly give AI models access to tools (code execution, web search, file I/O), the attack surface expands dramatically. Direct prompt injection on the model is one vector, but agents introduce new risks like tool misuse or multi-step manipulation.

For developers building agentic systems, practicing security in a sandboxed environment like GetJailbroken can reveal blind spots before deployment. The gamified structure—with distinct levels and a clean UI—makes it approachable for both beginners curious about prompt injection and experienced red-teamers looking to test advanced attack chains.