Llama Guard — agentic threat model

Llama Guard is an instruction-tuned LLM. It is highly vulnerable to adversarial jailbreaks, prompt injection bypasses, and model utility degradation if fine-tuned on poisoned safety datasets.

Not certain from the listing — The model relies on pre-defined safety taxonomies and training data. If users customize or fine-tune the model with custom datasets, data poisoning or lack of lineage tracking could introduce blind spots.

Not certain from the listing — Llama Guard acts as a utility rather than an orchestrator. Vulnerabilities would arise from how the hosting framework parses its binary classification outputs (e.g., failing to block a prompt due to parsing errors).

Not certain from the listing — Deployment is self-hosted or cloud-hosted. Risks include unauthorized API access to the model endpoint, denial of service (DoS) via resource-intensive inputs, and side-channel attacks.

Llama Guard is itself an observability and guardrail tool. Its primary threat is evaluation gaming, where attackers systematically probe the model to map its decision boundaries and find bypasses.

Not certain from the listing — While Llama Guard helps enforce safety policies (e.g., aligning with NIST AI RMF), the model itself does not enforce authentication, authorization, or audit logging out of the box.

Not certain from the listing — Llama Guard does not natively participate in multi-agent marketplaces, but a compromise or bypass of this guardrail could lead to cascading safety failures across downstream agents relying on it.