Skip to content

Mukund2/theron

Repository files navigation

Theron

Security proxy for agentic AI systems. Blocks prompt injection attacks automatically.

Why Theron?

Other tools scan for bad patterns. Theron enforces privilege separation.

Most AI security tools try to detect "malicious" prompts with regex or ML. The problem? Attackers constantly find bypasses.

Theron takes a different approach: content your AI reads should never have the same privilege as commands you issue. An email can't run shell commands - even if it contains valid-looking instructions.

What Makes It Different

Feature Theron Other Tools
Source Trust Tagging Tags where content came from (user vs email vs web) and enforces different permissions Treats all input the same
Zero Config curl | sh, restart terminal, done. Works with Claude Code, Cursor, etc. Requires SDK integration into your code
Local-First Runs entirely on your machine. No cloud, no data sent anywhere Most are SaaS or require cloud components
Honeypot Detection Seeds fake credentials - if your agent uses them, it's compromised Pattern matching only

Installation

curl -fsSL https://mukund2.github.io/theron/install.sh | sh
Windows (PowerShell)
irm https://mukund2.github.io/theron/install.ps1 | iex
Manual installation
pip install theron
theron setup

Restart your terminal. Your AI agents are now protected.

How It Works

┌─────────────────────────────────────────────────────────────┐
│                       YOUR COMPUTER                         │
│                                                             │
│  ┌─────────┐      ┌──────────┐      ┌───────────────────┐  │
│  │   AI    │ ───▶ │  THERON  │ ───▶ │ api.anthropic.com │  │
│  │  Agent  │ ◀─── │  Proxy   │ ◀─── │ api.openai.com    │  │
│  └─────────┘      └──────────┘      └───────────────────┘  │
└─────────────────────────────────────────────────────────────┘
  1. AI agent sends requests through Theron (automatic after setup)
  2. Theron tags each message with its source: USER_DIRECT, CONTENT_READ, TOOL_RESULT
  3. Tool calls are classified by risk: safe → moderate → sensitive → critical
  4. Policy matrix decides: user commands can run shell, but email content cannot
  5. Dangerous actions from untrusted sources are blocked automatically

The Core Idea: Source × Risk

Content Source Safe Tools Sensitive Tools Critical Tools
You typed it Allow Allow Allow
From an email/webpage Allow Block Block
From tool output Allow Block Block

This is why Theron stops attacks that pattern-matching misses. The attacker's payload might look legitimate, but it came from untrusted content - so it can't execute dangerous actions.

Features

  • Source Trust Tagging - Every message tagged with origin and trust level
  • Prompt Injection Detection - 50+ patterns across 6 attack categories
  • Risk-Based Gating - 4-tier tool classification with 100+ tools mapped
  • Honeypot Injection - Fake credentials seeded in outputs to detect compromise
  • Exfiltration Prevention - Blocks sensitive data from leaking to external services
  • Behavioral Baselines - Learns normal patterns, detects anomalies
  • Causal Chain Tracking - Visualizes attack path: email → parse → shell command

Dashboard

View protection status at http://localhost:8080:

  • Threats blocked
  • Requests checked
  • Recently blocked actions

CLI

theron setup              # Configure automatic protection
theron setup --status     # Check if protected
theron setup --uninstall  # Remove protection
theron                    # Start manually

Configuration

Config at ~/.theron/config.yaml:

proxy:
  listen_port: 8081

detection:
  sensitivity: 5           # 1-10
  injection_threshold: 70  # 0-100

gating:
  whitelist: []            # Always allow these tools
  blacklist: []            # Always block these tools

Security Model

Theron runs on localhost only - no authentication needed because only local processes can access it. DNS rebinding and CORS attacks are blocked.

License

MIT

About

Security layer for agentic AI - detects prompt injection, classifies action risk, blocks malicious commands

Resources

Stars

Watchers

Forks

Packages

 
 
 

Contributors