Sentinel AI Firewall
The security layer between your app and every LLM it calls.
Sentinel sits between your application and the LLM APIs it calls, scrubbing content for adversarial input on the way in and scanning tool outputs on the way back — catching what generic API gateways miss.
What it catches
Prompt injection & jailbreaks
Authority hijacks, prompt-extraction attempts, and persona-shift attacks are caught before they ever reach the model, whether they arrive as a direct user message or hidden inside a tool result.
PII & credential leaks
A dedicated secret detector scans content independently of threat scoring, redacting API keys, tokens, and credentials from Anthropic, OpenAI, Stripe, GitHub, AWS, and Slack before they leave your pipeline.
Poisoned RAG data
Retrieved documents and tool outputs are scanned with the same pipeline as user input, so adversarial content planted in a knowledge base or fetched page can't smuggle instructions into the model.
Agentic tool-call abuse
The agentic proxy scans tool_result content before it returns to the agent across Anthropic, Grok, OpenAI, and Gemini — with provenance-aware trust scoring so a compromised dependency never inherits a trusted directory's discount.
How it works
Normalize
Strips invisible characters, resolves homoglyphs, and applies NFKC normalization before anything is scanned — the original text is always preserved for the caller.
Fast-path match
A library of regex patterns catches high-confidence attack signatures with near-zero latency.
Deep-path similarity
Anything ambiguous is embedded and compared against a library of attack-signature embeddings using cosine similarity, with separate thresholds for standard and strict modes.
Providers
Four providers, one integration
Anthropic, Grok (xAI), OpenAI, and Google Gemini are all fully supported on the agentic proxy — point your existing SDK at Sentinel and nothing about your request/response handling changes.
See setup docs on sentinelaifirewall.com