RESK develops an open-source ecosystem for securing AI deployments at every layer: input-time attack detection, token-level output filtering, per-user RBAC, MCP security, semantic caching, and agent observability. Each project has a dedicated page — start there for features, quick starts, and demos.
The deployable LLM firewall: an OpenAI-compatible endpoint with RBAC capability bitmasks, editable filtering policies, multi-provider routing, and a full admin console. Live at demo.resk.fr.
Stops prompt injections before they reach your LLM — in one line. 11 attack-specific detectors, PII/canary defense, FastAPI middleware.
Per-user LLM firewall: every candidate token is checked inside the generation loop — unauthorized content becomes physically ungeneratable.
The shadow-ban logits engine: vectorized Aho-Corasick matching, streaming, vLLM adapter. Apache 2.0.
Security and management layer over the Model Context Protocol SDK: JWT auth, validation, PII detection, rate limiting, dashboard.
AI agent logger: every agent action recorded — functions, parameters, token probabilities — exported to Datadog, Prometheus, OTel.
Prompt-injection defense for Node & Bun — zero dependencies, 11 detectors, Express & Hono middleware. Live in-browser demo on its page.
Semantic LLM response caching with vector search: up to 90% API cost reduction, sub-5ms cached responses, OpenAI-compatible.
Deception for AI scanners: serves fake secrets and prompt-injection decoys at probed paths (/.env, /.aws/credentials) and fingerprints the attacker. Zero dependencies, Express & Hono middleware.
End-to-end encryption for AI conversations with policy-based, revocable access control.
All repositories are public on GitHub — star your favorites and contribute.