LlamaFirewall

 LlamaFirewall is an open-source guardrail framework from Meta, released in 2025, that serves as a real-time layer of defense for language model agents. It combines several guardrails into one policy engine: PromptGuard 2 detects jailbreak attempts, Agent Alignment Checks audit an agent's chain of thought for signs of goal hijacking, and CodeShield runs static analysis on generated code to catch insecure patterns. Developers extend it with custom regex and model-based scanners, and Meta runs LlamaFirewall in production, releasing it so teams can compose their own agent defenses.

Key Capabilities:

• PromptGuard 2 for real-time detection of jailbreak attempts

• Agent Alignment Checks that audit agent reasoning for goal hijacking

• CodeShield static analysis that flags insecure generated code

• A unified policy engine that composes multiple scanners into a pipeline

• Custom regex and model-based scanners for application-specific risks

• Layered defenses covering user input, agent reasoning, and code output