🛡️ AI Code Risk Firewall

Catches security, secret, and authorization risks in AI-generated code before you run, commit, or deploy. It watches the files you or your coding agent (Claude Code, Cursor, Copilot, Antigravity) modify and warns you in real-time right inside the editor.

Unlike heavy enterprise scanners, this is a sharp, change-focused firewall: fast, runs entirely local, and operates completely offline.

🚀 Key Features

• ⚡ Real-Time Change Watcher — Scans the active file as you type (debounced) and instantly on open or save.
• 🔒 Privacy-First & Offline — Runs 100% locally using regex/AST patterns. No code ever leaves your machine, and no API keys are required.
• 📊 Live Risk Score — Calculates a dynamic 0–100 risk score per file and displays it in the status bar (uses diminishing returns so multiple minor alerts don't outweigh a single critical severity issue).
• 📋 Interactive Risk Report — A rich side-panel webview grouping findings by severity with click-to-reveal jump-to-line navigation.
• 🤖 Agent Rules Generator — Automatically generates agent instructions (.cursorrules, .antigravityrules, CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md) derived from the active firewall rules to stop agents from writing vulnerable code in the first place.
• 📦 Dependency Diff Watcher — Inspects manifests and lock files (package.json, package-lock.json, yarn.lock, pnpm-lock.yaml) for typosquats of popular packages, known compromised versions, and suspicious lifecycle install scripts.
• 💡 One-Click Quick Fixes — Every finding gets lightbulb actions: silence a false positive on a single line, mute a noisy rule for the whole file, or jump straight to the full report — no manual comment typing.
• 🙈 Inline Suppressions — Mark intentional code with ordinary comments the firewall reads back in: risk-firewall-ignore-line, risk-firewall-ignore-next-line, or risk-firewall-ignore-file (optionally scoped to specific rule ids). Works in any language.
• 🔀 Scan Git Changes — One command scans exactly what you're about to commit — your staged changes (or working-tree changes when nothing's staged) — so AI-written diffs get a security gate before they land.

🔍 Security Rules & Detection Engine

The firewall categorizes and reports vulnerabilities across several dimensions:

1. Hardcoded Secrets (secret)

• AWS access key IDs (AKIA... / ASIA...) — [Critical]
• OpenAI API keys (sk-...) — [Critical]
• Anthropic API keys (sk-ant-...) — [Critical]
• Stripe secret/restricted keys (sk_live... / rk_test...) — [Critical]
• Google API keys (AIza...) — [High]
• GitHub tokens (ghp_... / gho_...) — [Critical]
• Slack tokens (xoxb-... / xoxp-...) — [High]
• Private Key blocks (BEGIN PGP/RSA/EC PRIVATE KEY) — [Critical]
• JSON Web Tokens (eyJ...) — [Medium]
• Generic assignments (api_key = "...", password = "...") — [High]
• Frontend exposure — Service role keys (SERVICE_ROLE) in frontend directories or secrets prefixed with NEXT_PUBLIC_ — [Critical/High]

2. Dependency Risk Guard (dependency)

• Typosquatting — Checks if added packages are single-edit-distance typos of the top 50 most-used packages (e.g. lodaash instead of lodash). — [Critical]
• Compromised Packages — Identifies packages tied to publicized supply-chain incidents (e.g. event-stream malicious versions, node-ipc protestware, sabotaged colors/faker). — [High]
• Lifecycle Install Scripts — Flags preinstall, install, or postinstall commands in package.json that invoke shell, network, base64 decoding, or execution scripts. — [High/Medium]

3. Attack & Injection Vectors

• Code Injection (injection) — Use of eval(), new Function(), shell commands with string interpolation (exec("...${var}")), and Python os.system() or shell=True subprocesses. — [High]
• SQL Injection (injection) — Assembled queries via interpolation or concatenation instead of parameterized queries. — [High]
• Cross-Site Scripting (xss) — Use of dangerouslySetInnerHTML, .innerHTML assignments, and document.write(). — [High/Medium]

4. Encryption & Networks

• Weak Cryptography (crypto) — Use of MD5 or SHA-1 for hashing, hardcoded JWT signing secrets, or Math.random() to generate security tokens. — [High/Medium]
• Disabled TLS (network) — Bypassing server certificate verification (rejectUnauthorized: false, python verify=False, or setting NODE_TLS_REJECT_UNAUTHORIZED=0). — [High]
• Insecure Endpoints (network) — Outbound remote requests utilizing plaintext http:// instead of https://. — [Low]

5. Config & Logging (config / auth)

• Debug Mode — Static debug switches (DEBUG = true) left active in production. — [Medium]
• Sensitive Logs — Logging or printing secret variables (console.log(password)). — [Medium/Low]
• Session Storage (auth) — Storing tokens/sessions in localStorage instead of Secure, httpOnly cookies. — [Medium]

Walkthrough

🛒 Installation & Marketplace

In the Editor (VS Code / Antigravity / Cursor / VSCodium)

Open the Extensions side panel in your editor and search for:

• martian7777.ai-code-risk-firewall (the unique extension ID)
• Or simply search: AI Code Risk Firewall

On the Web

You can also view and install the extension directly from the public registries:

• VS Code Marketplace: https://marketplace.visualstudio.com/items?itemName=martian7777.ai-code-risk-firewall
• Open VSX Registry: https://open-vsx.org/extension/martian7777/ai-code-risk-firewall

🛠️ Getting Started

Local Setup

1. Clone the repository and install the development dependencies:

   npm install
   

2. Build the extension:

   npm run build
   

3. Open the repository in VS Code or Antigravity and press F5 (or go to Run and Debug -> Run Extension).
4. An Extension Development Host window will open. Open the demo/vulnerable-sample.ts file in that window to see the real-time firewall diagnostics in action!

Commands List

Access these via the Command Palette (Ctrl+Shift+P or Cmd+Shift+P):

• Risk Firewall: Scan Current File — Run an immediate scan on the active editor and show report.
• Risk Firewall: Scan Whole Workspace — Scan all project files (ignoring excluded paths) and compile a project-wide report.
• Risk Firewall: Scan Git Changes (staged / uncommitted) — Scan only the files you're about to commit and report on them.
• Risk Firewall: Show Risk Report — View the live report panel side-by-side with your code.
• Risk Firewall: Generate Agent Security Rules — Create customized rules for AI coding assistants.

Suppressing False Positives

The firewall is built for recall, so it occasionally flags intentional code. Silence a finding with a comment — use your language's comment syntax (//, #, --, …):

const apiKey = process.env.OPENAI_KEY ?? "sk-localtestkeyonly"; // risk-firewall-ignore-line secret/openai-key

// risk-firewall-ignore-next-line
eval(trustedExpression);

• Append no rule id to silence every finding at that location, or list one or more ids/categories (comma- or space-separated) to scope it: risk-firewall-ignore-line secret, cors.
• Place risk-firewall-ignore-file anywhere in a file to mute matching rules for the whole file.
• The easiest way to add these is the lightbulb / Quick Fix menu (Ctrl+.) on any flagged line — it writes the comment for you.

⚙️ Configuration

You can customize the firewall behavior via your workspace settings.json:

📋 Release Notes

v0.1.1

Makes findings actionable — the firewall now helps you fix and dismiss risks, not just spot them.

1. One-Click Quick Fixes — Lightbulb code actions (Ctrl+.) on any flagged line: silence a single false positive, mute a noisy rule for the whole file, or jump to the full report. The suppression comment is written for you with the correct comment syntax and indentation.
2. Inline Suppressions — Honor risk-firewall-ignore-line, risk-firewall-ignore-next-line, and risk-firewall-ignore-file comments (optionally scoped to specific rule ids/categories) in any language. Suppressed findings drop out of diagnostics, the status-bar score, and the report.
3. Scan Git Changes — New Risk Firewall: Scan Git Changes command scans exactly what you're about to commit (staged changes, or working-tree changes when nothing is staged) via the built-in Git API — no extra dependencies.

v0.0.1 (Initial Release)

This is the first release of the AI Code Risk Firewall VS Code extension.

What's Implemented:

1. Security & Secrets Linting Engine — 30+ regex-based static analysis rules running locally in Node.js to capture high-severity leaks and common vulnerabilities.
2. Local Dependency Diff Watcher — Offline manifest analysis in package.json and lock files checking for typosquats, flagged protestware/malware packages, and automatic install hooks.
3. Interactive Webview Risk Report — A customized panel summarizing workspace/file health with a custom 0-100 risk score and direct code navigation links.
4. Agent Rules Generator — A generator targeting .cursorrules, .antigravityrules, CLAUDE.md, AGENTS.md, and .github/copilot-instructions.md to instruct AI models on security patterns automatically.
5. Debounced Live Watcher — Seamless editor integrations utilizing diagnostics markers without slowing down typing or compiler performance.

📄 License

This project is licensed under the MIT License.