gendigitalinc/sage

Sage – Safety layer for AI coding assistants

What it is – Sage is a lightweight security plug‑in that sits between an AI‑powered coding assistant (Claude Code, Cursor, VS Code, OpenClaw, OpenCode, etc.) and the host system. Whenever the assistant tries to run a tool call – e.g. a shell command, download a URL, write a file, or install a package – Sage intercepts the request, runs a series of threat checks, and only lets the action proceed if it is deemed safe.

Why it matters – Large language model (LLM)‑based coding tools can be tricked into executing malicious commands or fetching harmful payloads (prompt‑injection attacks, supply‑chain compromises, etc.). Sage adds a defensive “firewall” that protects the developer’s machine without requiring changes to the underlying AI model.

---

Core capabilities

Feature	What it does
URL reputation	Queries a cloud service to flag known malware, phishing or scam links before they are fetched.
Local heuristics	Over 300 YAML‑defined patterns detect dangerous shell commands, suspicious URLs, credential leaks, and code obfuscation.
Prompt‑injection detection	A two‑stage guard – fast heuristic rules followed by a fine‑tuned machine‑learning model – that spots injected instructions in fetched content.
Package supply‑chain checks	Looks up npm / PyPI packages, verifies they exist, evaluates their reputation and age to catch typosquatting or abandoned libraries.
Plugin scanning	At the start of each AI‑assistant session, Sage scans any installed plug‑ins for known threats.
AMSI integration	On Windows (and WSL) Sage forwards commands to the Antimalware Scan Interface, letting the OS’s real‑time scanner block malware.

How to get it

1. Read the online install guide – https://ai.gendigital.com/sage (keeps the steps up‑to‑date).
2. Choose your AI assistant and follow the short command shown in the README:
   • Claude Code – add the repo as a plugin and install sage@sage.
   • Cursor – install the Gen Sage extension from the Cursor marketplace.
   • VS Code – install the Gen Sage extension from the VS Code marketplace.
   • OpenClaw – openclaw plugins install @gendigital/sage-openclaw.
   • OpenCode – add @gendigital/sage-opencode to the opencode.json config file.
3. After installation, Sage runs automatically, logging decisions to an on‑disk JSONL audit file.

Typical workflow

1. The developer writes a prompt to the AI assistant.
2. The assistant generates a tool call (e.g., git clone https://evil.example.com/repo.git).
3. Sage intercepts the call and:
   • Checks the URL against the cloud reputation service.
   • Runs the local YAML heuristics.
   • (If needed) feeds the content to the prompt‑injection ML model.
   • Looks up any package involved in a supply‑chain check.
   • Sends the command through AMSI on Windows.
4. If any check flags a threat, Sage blocks the action and records the event; otherwise the command is allowed to execute.

Who should use it

• Developers who rely on LLM‑based code assistants and want to avoid accidental malware execution.
• Security teams looking for a plug‑and‑play safeguard that works across multiple AI IDE extensions.
• Ops / DevOps who automate code generation and need a policy‑enforced gate before any generated script runs.

Extensibility & contribution

• Threat patterns live in threats/ as simple YAML files – contributors can add new rules via a pull request.
• The repository includes a Developer Guide describing the architecture, testing harness, and how to train or update the prompt‑injection ML model.
• All code is Apache‑2.0; the rule set has its own Detection Rule License 1.1.

---

Bottom line – Sage provides a focused, cross‑assistant security layer that catches malicious tool calls generated by AI coding assistants, using a mix of cloud reputation, local heuristics, ML‑based prompt‑injection detection, and OS‑level scanning.