Glossary / Latent Defense

Definition

What is Latent Defense?

Embedding-layer security

Latent defense is AI security for the embedding layer — the vector geometry your retrieval store searches and your agent acts through. Trinitite runs six on-by-default, fail-open defenses against RAG poisoning and prompt injection, including an embedding-based Agent Action Guard that judges the proposed action’s semantics instead of the agent’s hijackable reasoning.

Attackers stopped fighting your prompt and started fighting your math — reshaping the vectors behind retrieval, the query, the action, and the policy clause a verdict rests on. The six defenses: hybrid keyword+semantic retrieval that defeats gradient-guided RAG poisoning; black-hole (hubness) detection that quarantines retrieval-magnet vectors; covariance-aware per-cluster Mahalanobis scoring; query-side manifold scoring for an adversarial-probe signal;

the Agent Action Guard, an independent embedding gate that survives prompt injection because it scores the tool call’s semantics, not the justification; and policy-clause anchoring that binds every verdict to the exact governing clause in the tamper-evident chain. All six reuse the embedding and vector-store seams you already run — no new infrastructure — and crosswalk to EU AI Act Art. 9–17 and GDPR Art. 22.