LlamaFirewall layers PromptGuard 2 for jailbreak detection, Agent Alignment Checks that audit an agent's reasoning for injection and goal drift, and CodeShield for static analysis of generated code. The authors call the alignment check experimental. Learn what each layer inspects and where it sits in an agent loop.
Read in this order
- OfficialLlamaFirewall: an open source guardrail system (arXiv)The paper describing the three components and their intended role as a final defense layer.