# AITHURA — Full Reference

> AITHURA is one control layer for every AI action. It is a zero-trust, BYOK security gateway for LLM calls, AI agents, MCP servers, and tool use. It secures AI activity, governs agent actions, and proves what happened.

Website: https://aithura.com
Docs: https://aithura.com/docs
Source: https://github.com/AITHURA-com/aithura

## What AITHURA is

AITHURA is an inline security and governance layer. Applications send LLM requests and agent tool calls to AITHURA instead of directly to providers. AITHURA scans each request with the Sentinel Protocol engines, enforces policy, forwards allowed traffic to the upstream provider using the customer's own key, and records evidence.

AITHURA is not an LLM provider and does not resell model access. It is BYOK: customer provider keys (OpenAI, Anthropic, Google, or any OpenAI-compatible endpoint) are stored encrypted with AES-GCM-256.

## Who it is for

- Developers shipping LLM features or agents who need prompt-injection and data-leak protection without rewriting code.
- Enterprises that need AI governance, agent access control, and audit evidence for SOC 2, NIST AI RMF, and EU AI Act programs.
- Government, defense, and defense-industrial-base teams that need air-gapped deployment and evidence for NDAA 1513, NIST SP 800-53/800-171, DFARS 252.204-7012, CMMC Level 2, and the DoD IL4/IL5 path.

## Three outcomes

1. Secure — block prompt injection, jailbreaks, data exfiltration, PII and secret leakage, and unsafe output.
2. Govern — control which tools agents can call with explicit allow / warn / block modes; explicit modes override automated risk scores; unknown tools are auto-added in warn mode.
3. Prove — signed action receipts, AI bill of materials (AIBOM), forensic replay, and exportable compliance audit logs.

## Sentinel Protocol

- 81 deterministic security and governance engines (no model-in-the-loop for core decisions, so results are repeatable).
- Open source; runs as a local sidecar on port 8787 or inside AITHURA's hosted gateway.
- Engine families include: prompt injection detection, PII and secret detection, data exfiltration detection, dangerous shell / command detection, tool schema validation, capability scoping, MCP shadow-server and poisoning detection, memory integrity, canary tool traps, cross-tenant isolation, lateral movement defense, and output safety checks.
- Scanning happens before forwarding: block-eligible or high-confidence injection prompts receive a "blocked" error and never reach the provider.

## Integration (LLM gateway)

OpenAI-compatible. Change the base URL to the AITHURA gateway and authenticate with an AITHURA API key (prefix `aithura_live_`), sent either as an `x-aithura-key` header or `Authorization: Bearer <key>`. Save your provider key once in the dashboard (Model Providers). See https://aithura.com/docs for the exact endpoint.

Python example:

    from openai import OpenAI
    client = OpenAI(base_url="<AITHURA gateway URL from docs>", api_key="aithura_live_...")
    client.chat.completions.create(model="gpt-4o-mini", messages=[{"role": "user", "content": "Hello"}])

## Integration (Agent Shield)

Agent tool calls are sent to the Agent Shield endpoint with the AITHURA key. Upstream tool URLs are configured per tool in the dashboard (Agent Tools page); per-request upstream URLs are ignored for security (SSRF protection). Blocked tools return HTTP 403. Every decision appears in Agent Events.

## Deployment options

- Hosted API (fastest start).
- Private VPC.
- Self-hosted Sentinel sidecar (Docker, port 8787).
- Fully air-gapped / DDIL environments.

## Data handling

- Provider keys encrypted with AES-GCM-256.
- AITHURA API keys stored as SHA-256 hashes.
- Stored prompts and responses truncated to 100 characters; compliance CSV exports truncated to 200 characters.
- Row-level security on all customer data.

## Compliance and framework mapping

OWASP LLM Top 10 (LLM01–LLM10), NIST AI RMF, SOC 2 evidence support, EU AI Act traceability readiness, NDAA Section 1513, NIST SP 800-53, NIST SP 800-171, DFARS 252.204-7012, CMMC Level 2, DoD IL4/IL5 path. AITHURA provides evidence and controls that support these programs; it does not by itself certify compliance.

## Frequently asked

- What is AITHURA? A BYOK zero-trust gateway that secures, governs, and records every LLM call and agent action.
- Does AITHURA host models? No. You bring your own provider keys.
- How does it stop prompt injection? Deterministic Sentinel engines scan each prompt before forwarding and block high-confidence attacks.
- Can it run offline? Yes — the Sentinel sidecar can be self-hosted and air-gapped.
- What is an AIBOM? An AI bill of materials: an inventory of models, tools, and policies involved in AI actions, used as audit evidence.