Aegis is building security for the agentic eraV1.0 Release Candidate → GitHub

Product

A runtime security gateway for AI agents.

Aegis sits between autonomous agents and the tools they use. Every action is evaluated before it executes — so prompt injections, rogue agents and malicious tool responses cannot cause unauthorized harm.

01

Intercept

Agents call Aegis before executing any tool action — via the Python SDK, REST API or the MCP Security Gateway.

02

Evaluate

Permission, trust, policy, risk and threat engines run independently of the LLM, with strict precedence and fail-closed behavior.

03

Enforce

ALLOW executes, BLOCK stops the action, REVIEW pauses it for a human decision — every outcome lands in the audit trail.

The pipeline

Independent. Deterministic. Explainable.

Security evaluation is decoupled from the LLM entirely. Decisions are reproducible, reasons are recorded, and failures fail closed.

See how it works

  1. ACTION

    Agent intent is intercepted

  2. CONTEXT

    Agent, tool and session verified

  3. THREAT ANALYSIS

    Deterministic detectors run

  4. RISK SCORE

    Explainable 0–100 assessment

  5. POLICY

    Rules evaluated with precedence

  6. DECISION

    ALLOW · REVIEW · BLOCK

Defenses

Four defenses on every action.

Prompt Injection Defense

01

Deterministic detectors and optional AI security intelligence analyze untrusted content and tool outputs for instruction-override attempts.

Tool Call Protection

02

Dangerous operations, unauthorized tools and broad-scope targets are stopped before execution.

Sensitive Data Protection

03

Sensitive parameters are redacted before logging — keys, tokens and credentials never enter the audit trail in plaintext.

Unauthorized Action Detection

04

Agent identity, permissions, trust classification and user delegation are verified on every single call.

Components

One platform, five surfaces.

Python SDK

aegis_sdk — drop-in client for Python agents with typed errors for BLOCK and REVIEW outcomes.

MCP Security Gateway

Zero-trust interception for any Model Context Protocol server — no server-side changes required.

VS Code Extension

Approvals, agent status and attack scenarios from inside the IDE.

Web Command Center

Approvals, audit exploration, policies and threat monitoring in one console.

Attack Lab

Built-in adversarial scenarios to validate controls against prompt injection, exfiltration and destructive actions.

Docker deployment

Production-ready docker-compose with PostgreSQL and Redis.