Open Source

GitHub Repository

Hugging Face Collections

Patronus Ark

The open-source, on-device AI security library for prompt injection, DLP, PII and agentic tool risks. A small Rust core with first-class Python bindings keeps sensitive traffic local, fast and auditable.

View the repository

Read the library docs

patronus-ark / security gateway

policy ready

pip install patronus-ark

SecurityGateway(categories=[injection, dlp, pii])

L1–L3

layered scanning

Rust + Python

core + bindings

Layered, local-first security scanning

One security gateway for every prompt, response and tool call.

Patronus Ark inspects AI traffic locally, without sending prompts, documents or tool outputs to a cloud service. Native rules handle most traffic in microseconds; uncertain cases can escalate through cached model layers.

Apps + agents

Chat, copilots, RAG services and autonomous tools.

Prompt gateway

PII, secrets, jailbreak and policy checks before inference.

Security orchestrator

Routes each event through pipelines, models and policy actions.

HF model checks

Adapters for safety classifiers, embeddings and risk scoring.

Enforce + audit

Allow, block, redact, route, log and prove control coverage.

Risk scoring across prompts, responses and tools

Policy actions: allow, warn, redact, block, escalate

Structured audit events for AI governance and SOC review

L1 / L2 / L3 layered scanning

Security coverage for prompts, data and agent actions.

Use one local security gateway for prompt injection, DLP, PII, sensitive documents, tool-call risks, routing and threat detection.

01 Prompt injection

Detect jailbreaks and instruction overrides before they reach a model or agent.

02 DLP + secrets

Catch credentials, sensitive data and policy violations in prompts and responses.

03 Tool-call controls

Approve, deny or transform agent actions before tools touch files, APIs or internal systems.

04 PII + documents

Inspect retrieval context and document chunks for personal data, leakage and injection payloads.

05 Model routing

Route uncertain cases to the right cached model layer without giving up local control.

06 Tool classes + actions

Classify what an agent wants to do and enforce policy before execution.

07 Threat detection

Score malicious intent and suspicious activity across applications and workflows.

08 Offline + air-gapped

Keep scanning available where connectivity, privacy or data sovereignty is non-negotiable.

09 Audit-ready results

Emit structured results for governance reviews, security operations and commercial deployments.

Benchmark-ready, resource-aware scanning

Stable below 1GB RAM with approximately 1mb/s throughput. Measure security decisions independently from model latency and tune the gateway for your own workload.

<1GB

RAM stable

~1mb/s

throughput

Models on Hugging Face

Explore the Patronus model collections for prompt injection, PII, secrets and safety classification. Keep inference local, self-hosted or routed through your approved provider.

Browse Patronus model collections →

Ship Patronus Ark under your own license.

Use GPL-3.0-only for open-source products, or talk to us about a commercial license for proprietary deployments. We can help your team bring local scanning, policy enforcement and audit-ready results into production.

Star the repository

Request an Enterprise License