CloneGuard

Hook-level prompt injection defense for AI coding agents.

Your AI agent reads untrusted repos. CloneGuard watches what it does next.

CloneGuard runs at the hook layer -- before tool calls execute, outside the agent's control. It detects prompt injection attempts, constrains suspicious operations via OS-level sandboxing, and emits structured audit logs.

• 5-Minute Setup

  Install from PyPI, run one command, and CloneGuard is active in Claude Code.

  Getting started

• 240 Detection Rules

  Pattern matching, semantic classification, and behavioral sequence monitoring across 34 attack categories.

  Detection engine

• Detect, Constrain, Audit

  Three-verdict enforcement with OS-level sandboxing. Dry-run by default.

  Enforcement

• Works With Any Agent

  Built for Claude Code. Standalone scan works with any agent. Hook protocol compatible with Gemini CLI, Cursor, and Windsurf.

  Getting started

Install

pip install cloneguard            # Pattern matching only
pip install "cloneguard[mini]"    # + semantic classifier (recommended)

Development Status

CloneGuard is in active development (v0.6.0). The core detection engine is tested with 1,677 automated tests and false positive rates calibrated against 208,127 real coding-agent sessions from published SWE-bench datasets.

Enterprise features (policy backends, SIEM connectors, fleet deployment) are early-stage and experimental.

We want feedback -- open an issue or contribute.