AI Guardrails

Every message into an AI agent and every reply out of it gets one verdict and the name of the rule that decided. Thirteen checks run in a fixed order against a JSON policy: Unicode canonicalisation, container-breakout patterns, prompt injection, blocklists, length limits, harmful content, PII and credentials, data exfiltration, and system-prompt leakage. An optional NVIDIA Nemotron model on Nebius reviews only what the rules allowed and can escalate to a block but never release one, so a broken or manipulated model degrades to rules-only instead of no security. A default-deny gateway adds per-user TOTP two-factor, scoped API keys, CSRF checks, rate limiting and a hash-chained audit log. Zero runtime dependencies; built with Isabella Wu and Annie Liao.

October 2026