LlamaFirewall: an open-source guardrail system for building secure AI agents

The paper's authors propose LlamaFirewall, an open-source guardrail framework for building secure AI agents, as a last line of defense for AI agent risk protection. It contains three guardrails: PromptGuard 2 is a general-purpose jailbreak detector that the authors claim achieves clearly leading performance; Agent Alignment Checks inspects the agent's chain of thought to detect prompt injection and goal deviation, and remains experimental; CodeShield is a fast, scalable online static analysis engine for blocking unsafe or dangerous code generated by coding agents. The framework also provides customizable scanners, and developers can update guardrails using regular expressions or LLM prompts.