aegisflow 2.2.0

A sophisticated Security Liaison and Governance Layer for AI Agents

AegisFlow v2.2.0 - Governance Layer for AI Agents

AegisFlow is a sophisticated Security Liaison designed to govern AI agent actions through transparent mediation rather than silent blocking. It acts as a "conscious" layer, ensuring high-risk operations are verified by a human-in-the-loop (HITL).

Core Philosophy

• Suspicion Scoring: Every action is assigned a Threat Level (Low, Medium, High).
• Transparent Mediation: Risks are reported clearly; high risks require explicit approval.
• Sentinel State Engine: Tracks reputation and persists logs.
• Audit Trail: All decisions and outcomes are logged to ~/.aegis/logs/aegis_audit.json.
• Sandwich Wrapper: Wrap any terminal command in a monitored shell.

Installation

pip install aegisflow

This installs the aegis CLI tool globally.

Usage

1. The AegisSandwich (Universal Terminal Wrapper)

Run aegis run to wrap any agent process. AegisFlow will monitor its output for dangerous patterns and suspend it if necessary.

aegis run python my_agent.py

2. Static Scan

Scan a file for behavioral redlines:

aegis scan path/to/script.py

3. Universal LLM Integration (Code)

Wrap any LLM call with SafeGenerator to get instant security:

from aegisflow.llm import SafeGenerator

# Automatically scrubs keys, checks for injections, and verifies dangerous outputs.
llm = SafeGenerator()

response = llm.generate("Write a script to delete all files.", model="gpt-4")
print(response)

Sentinel State Engine

The Sentinel tracks "Risk Streaks". If an agent triggers 3 Medium risks in a row, the next action is automatically escalated to High.

For High Risk (or escalated) actions, the user must provide a Reasoning String (e.g., "Debugging local server") to proceed. Simple "yes/no" confirmations are not accepted for high-risk operations.

Configuration (.aegis.json)

Create a .aegis.json in your project root or home directory to customize behavior:

{
  "protected_paths": [
    "/prod/db",
    "./secrets"
  ],
  "strict_mode": true
}

Behavioral Redlines

AegisFlow monitors for:

• Recursive Operations: rm -rf, massive deletes.
• Exfiltration: POST requests containing key-like patterns.
• Rule Negation: AI thoughts attempting to bypass security constraints.

License

MIT