Highflame Agent Control Platform

Highflame — security infrastructure for AI agents. Runtime guardrails, Cedar policy enforcement, agent identity, MCP gateway security, and adversarial red teaming.

Highflame is the security control plane for AI agents and autonomous systems. It provides runtime guardrails, typed Cedar policy enforcement, agent identity, MCP gateway security, model supply chain scanning, and adversarial red teaming — across a unified detection and observability layer.

Why Highflame

AI agents call tools, connect to MCP servers, modify code, invoke APIs, and operate across long-running workflows. Every one of those actions is a security decision. Highflame gives engineering and security teams the infrastructure to make agent autonomy safe to deploy:

• Sub-10ms inline guardrails — tiered detection (deterministic → ML → cloud) with Cedar policy evaluation on every prompt, tool call, and model response

• Tiered detection — fast & slow guardrail detection tiers for deterministic latency when applying guardrails & controls.

• Cedar-native policy — declarative, auditable authorization rules scoped to agents, tools, environments, and trust levels

• Agent identity (ZeroID) — OAuth2 tokens for non-human identities with delegation chains, scope enforcement, and credential policies

• Session-aware enforcement — cross-turn risk tracking with stateful ML detectors (GRU-based multi-turn jailbreak detection)

• Multi-product coverage — same policy and detection layer across Guardrails, Code Agents, MCP Gateway, Browser Security, and Red Team

Get Started

Quickstart

First guarded request in under 5 minutes

Securing Agents

Choose the right integration pattern: SDK, Gateway, or Trace ingestion

SDK Quick Reference

Python and TypeScript side by side — all common patterns

By Role

Security Teams

Provider setup, route creation, guardrail policies, governance

Red Team Testers

Adversarial testing of agents, prompts, and tool workflows

Developers

SDK integration, Shield wrappers, framework hooks

API & SDK

Python SDK

Decorators, async, LangGraph / CrewAI / Strands integrations

TypeScript SDK

Shield wrappers, streaming, zero runtime dependencies

REST API

POST /v1/guard, /v1/detect, /v1/guard/stream — curl examples

Platform

Guardrails & Policies

Threat catalog, Cedar cookbook, evaluation lifecycle

Agent Identity (ZeroID)

OAuth2 for agents — tokens, delegation, credential policies

MCP Gateway

Secure MCP server connections, tool governance, credential modes

Code Agents

IDE and coding assistant security — Cursor and Claude Code

Red Teaming

Adversarial testing, attack engines, model supply chain scanning

Observability

Traces, threat alerts, governance, OpenTelemetry integration

Production & Operations

Guide	Description
Production Patterns	Singleton clients, rollout stages (monitor → alert → enforce), graceful degradation
Multi-Tenancy	Per-tenant isolation, session scoping, Cedar context injection
Testing	Mock mode, response inspection, CI/CD integration
Troubleshooting	Common errors and resolution paths
Deployment	AWS, Azure, GCP deployment guides

Framework Coverage

Highflame maps detections, policies, and findings to industry security frameworks:

Framework	Coverage
OWASP LLM Top 10	Prompt injection, data leakage, tool misuse, model DoS
OWASP MCP Top 10	Tool poisoning, server impersonation, cross-origin escalation
MITRE ATLAS	Adversarial ML techniques mapped to detection signals
NIST AI RMF	Risk measurement, governance controls, audit evidence

Connect

• GitHub — SDKs, examples, and open-source projects

• Highflame — Product updates and company information

• Changelog — Platform updates, new detectors, and policy features