Agent Immune

This server brings adaptive security to AI agents through prompt injection detection, semantic memory, and output scanning. It exposes five MCP tools: assess_input evaluates user prompts for injection attempts, assess_output scans responses for credentials and PII, learn_threat adds new attack patterns to semantic memory, harden_prompt wraps system messages with role-lock guards, and get_metrics returns telemetry. The semantic layer catches paraphrased attacks that slip past regex rules by maintaining embeddings of confirmed threats. It ships with 50 curated attack patterns across 11 languages for instant protection, supports circuit breaking for sessions that hit block thresholds, and logs structured JSON events for observability. Reach for this when you need runtime defense that learns from incidents rather than just static rules, especially in agents handling sensitive data or untrusted user input.