What is an AI firewall?

An AI firewall is a security layer that inspects prompts, model responses, and agent tool calls before they reach your LLM or downstream systems. ArgusAI acts as an OpenAI-compatible proxy that blocks prompt injection, redacts PII, and enforces policy on every request.

How does ArgusAI protect against prompt injection?

ArgusAI runs a multi-layer detection pipeline: L1 pattern rules catch known injection signatures in milliseconds; L2 heuristics score obfuscated and paraphrased attacks; optional L3 LLM reasoning handles ambiguous cases. The engine normalizes Base64, hex, markdown, homoglyphs, and spacing evasions before matching.

Is ArgusAI compatible with OpenAI SDKs?

Yes. Point your OpenAI-compatible client at ArgusAI's /v1 base URL and use your ArgusAI project API key. No SDK changes are required for chat completions.

Does ArgusAI support MCP and agentic workflows?

Yes. Chat traffic goes through the proxy. For MCP security, call ArgusAI inspect endpoints for tool arguments and outputs. Growth plan and above adds cross-tool taint tracking when you pass session IDs.

What PII does ArgusAI detect?

ArgusAI detects emails, phone numbers, government IDs (including Aadhaar, PAN, SSN), credit card numbers, and IP addresses. You can configure block, flag, or redact behavior per project.

How do I try ArgusAI before integrating?

Use the free Playground at /playground to paste prompts and see layer-by-layer detection results — block, flag, redact, or allow — without writing code.

Yes. ArgusAI is in public beta — the Playground and docs are live now. Join the waitlist on the homepage for early access to the full gateway and dashboard onboarding.

AI Security Gateway

Public beta

AI Firewall for the Agentic Era

Chat completions: point your OpenAI-compatible client at ArgusAI — swap the base URL, keep your SDK. MCP security and taint tracking: add lightweight API calls to our inspect endpoints (Growth plan and above).

Try the Playground Join the waitlist

Diagram showing AI agent requests passing through the ArgusAI gateway shield before reaching LLM providers and MCP tool servers

Public beta

Now in public beta

Try prompt injection detection in the Playground today. Full gateway access, dashboard onboarding, and MCP security are rolling out to waitlist members first.

Available now

Interactive Playground with L1–L3 detection pipeline
Public docs, glossary, and security research
OpenAI-compatible proxy integration guide

Coming soon

Self-serve dashboard onboarding for all waitlist members
Paid plans and subscription management
MCP inspect API and taint tracking for Growth tier

Try the Playground Join the waitlist

Early access

Join the waitlist

Get notified when the full gateway, dashboard, and MCP security ship.

Agentic AI has an unguarded attack surface

Prompt injection slips through

Instructions hidden in user input, RAG documents, or tool output can hijack an agent's behavior mid-session.

PII leaks both directions

Sensitive data can enter a prompt or come back in a model response with no inspection layer in between.

MCP tool calls run unchecked

Tool arguments and tool output are rarely validated before an agent acts on them or passes them downstream.

No audit trail when something goes wrong

Without call-level logging, tracing a bad outcome back to its source after the fact is close to impossible.

One gateway, full visibility

ArgusAI sits between your application and any LLM provider as a transparent, OpenAI-compatible proxy. Every request gets a full threat scan (L1–L3 + PII) before it reaches the model; responses are scanned for PII and redacted when your policy requires it. MCP tool inspection is available via separate API endpoints you wire into your agent loop.

Capabilities

Defense in depth for every call

Pattern-based filtering

Catches direct prompt injection, jailbreak attempts, and credential leakage in milliseconds, before a request goes anywhere.

Heuristic escalation scoring (L2)

Boosts risk scores for paraphrased injections, escalation phrases, and obfuscated payloads that L1 patterns alone can miss. ML classifier path planned.

LLM reasoning for edge cases (L3)

Ambiguous prompts (L2 score 0.4–0.75) can get a Claude Haiku second opinion — Starter plan and above, requires Anthropic API key on the gateway.

Cross-tool taint tracking (L4)

Tracks data provenance across MCP inspect calls when you pass session IDs — Growth plan and above, requires Redis. Your agent calls our inspect endpoints; taint is not automatic on chat-only proxy traffic.

MCP tool call inspection

Call /v1/mcp/inspect-call and /v1/mcp/inspect-output from your agent to validate tool arguments and output against policy — Growth plan and above, explicit integration required.

PII detection and redaction

Identifies emails, phone numbers, government IDs (Aadhaar, PAN, SSN), and card numbers on ingress; redacts PII on egress responses when enabled.

Configurable policies

Set block, flag, or redact behavior per project — tune sensitivity without changing code.

Audit log

Calls are recorded with detection scores, action taken, and latency — persisted to MongoDB when connected, queryable per project. Retention varies by plan tier.

Why ArgusAI

Security for the full agent loop — not just the prompt

Most AI security tools focus on one slice: chat input, observability after the fact, or generic API filtering. ArgusAI is a gateway built for agentic systems — chat, MCP tools, and cross-call provenance in one place.

In-path enforcement, not post-hoc alerts

Threats are scored and acted on before requests reach your LLM or before tool output returns to your agent — block, flag, or redact per project policy.

Chat and MCP in one gateway

Point chat at our OpenAI-compatible proxy. Wire MCP inspect calls into your agent loop for tool arguments and tool output — the layers most scanners never see.

Layered detection that scales with risk

Millisecond pattern rules first, heuristic escalation second, LLM reasoning for edge cases when you need it — so you are not paying model latency on every call.

Drop-in integration

Keep your existing OpenAI SDK. Swap the base URL for chat; add lightweight inspect API calls where agents use tools. No full application rewrite.

Policy you control

Per-project settings for injection blocking, PII redaction, L3 thresholds, MCP tool allowlists, and token budgets — editable in the dashboard without shipping new code.

Audit without polluting context

Structured call logs with detection scores and actions taken — for incident response and compliance — kept out of model prompts.

MCP security and taint tracking require explicit integration and a Growth plan or above. Outbound webhooks, SIEM sink, and audit export require Pro+. L3 reasoning requires Starter+. See Plans for feature gates.

See integration docs →

Works with what you already have

OpenAI-compatible chat proxy — change a base URL, not your SDK
Supports OpenAI, Anthropic, and Gemini upstream providers
MCP inspect API for tool calls and tool output (Growth+; wire into your agent)
Per-project policy controls with plan-based feature gating

Live in three steps

Step 1
Get an API key
Sign in, complete onboarding, and create a project — your ArgusAI key is shown once in the dashboard.
Step 2
Point your SDK at /v1
Swap your OpenAI-compatible base URL for ArgusAI's proxy. Add MCP inspect calls separately if you need tool-level security.
Step 3
Monitor in the dashboard
Review calls, threat breakdowns, and analytics per project. Refresh to see new events as your agents run.

Enterprise

Custom Rego policies via OPA

Optional layer for regulated workloads: attach custom Open Policy Agent rules after the detection pipeline. Requires an OPA sidecar and gateway OPA_URL — not included in self-serve plans.

OPA setup guide

Frequently asked questions

What is an AI firewall?: An AI firewall is a security layer that inspects prompts, model responses, and agent tool calls before they reach your LLM or downstream systems. ArgusAI acts as an OpenAI-compatible proxy that blocks prompt injection, redacts PII, and enforces policy on every request.
How does ArgusAI protect against prompt injection?: ArgusAI runs a multi-layer detection pipeline: L1 pattern rules catch known injection signatures in milliseconds; L2 heuristics score obfuscated and paraphrased attacks; optional L3 LLM reasoning handles ambiguous cases. The engine normalizes Base64, hex, markdown, homoglyphs, and spacing evasions before matching.
Is ArgusAI compatible with OpenAI SDKs?: Yes. Point your OpenAI-compatible client at ArgusAI's /v1 base URL and use your ArgusAI project API key. No SDK changes are required for chat completions.
Does ArgusAI support MCP and agentic workflows?: Yes. Chat traffic goes through the proxy. For MCP security, call ArgusAI inspect endpoints for tool arguments and outputs. Growth plan and above adds cross-tool taint tracking when you pass session IDs.
What PII does ArgusAI detect?: ArgusAI detects emails, phone numbers, government IDs (including Aadhaar, PAN, SSN), credit card numbers, and IP addresses. You can configure block, flag, or redact behavior per project.
How do I try ArgusAI before integrating?: Use the free Playground at /playground to paste prompts and see layer-by-layer detection results — block, flag, redact, or allow — without writing code.
Is ArgusAI in beta?: Yes. ArgusAI is in public beta — the Playground and docs are live now. Join the waitlist on the homepage for early access to the full gateway and dashboard onboarding.

Built with security in mind

Ingress: full threat pipeline plus PII scan; egress: PII detect and redact when policy allows
Audit logs are structured data, never re-injected into model context
Policy is configurable per project — start permissive, tighten over time

Read our privacy policy