🛡️ SecureVibes

AI‑Native Security for Vibecoded Apps

SecureVibes uses Claude’s multi‑agent architecture to find security issues with concrete evidence. Agents coordinate to map your architecture, model threats, review code, and generate a clear report. An optional DAST phase validates exploitability via HTTP testing using auto‑discovered skills.

---

🚀 Quick Start

# Install for the latest release on PyPi (might not have all the latest changes in the code)
pip install securevibes

# NOTE: the package uploaded on PyPi might not have all the latest changes. 
# I will try to release a new version of the package whenever there are significant changes/developments
# If you would rather use the version with the latest changes, you can do the following:

# Install for the latest version (might be buggy)
git clone https://github.com/anshumanbh/securevibes.git
cd securevibes
virtualenv env
. env/bin/activate
pip install -e packages/core

# Authenticate (choose one method)
# Method 1: Session-based (recommended)
# You could use your Claude subscription here, if you don't want to pay per API requests
claude  # Run interactive CLI, then type: /login

# Method 2: API key
export ANTHROPIC_API_KEY="your-api-key-here"

# Scan your project
securevibes scan /path/to/code --debug

# The most important part
# Sit back and relax. Please be patient as the scans might take some time, depending upon the model being used.

Get your API key from: https://console.anthropic.com/

---

🔐 Runtime Safety Model

SecureVibes runs scan orchestration with Claude SDK permission_mode="default". Runtime safety depends on explicit tool surfaces and scanner hooks (including repository-boundary guardrails, PR review artifact constraints, and DAST database CLI blocking).

Operational guidance:

• Run scans only on trusted repositories.
• Prefer isolated CI runners/containers for scans.
• Avoid running scans with access to production credentials or sensitive networks.

---

🤖 Agents

• Assessment → SECURITY.md
• Threat Modeling (STRIDE + technology-specific skills) → THREAT_MODEL.json
• Code Review → VULNERABILITIES.json
• Report Generator → scan_results.json
• DAST (optional) → DAST_VALIDATION.json (validates via HTTP when --target-url provided)

🤖 Agentic Detection

SecureVibes automatically detects agentic applications (LLM APIs, agent frameworks, tool execution) and requires OWASP ASI threats in threat models. Override with --agentic or --no-agentic.

🌍 Supported Languages

11 Languages: Python, JavaScript, TypeScript, Go, Ruby, Java, PHP, C#, Rust, Kotlin, Swift

Smart Features:

• Auto-detects languages in your project
• Language-aware exclusions (Python: venv/, JS: node_modules/, Go: vendor/)
• Handles mixed-language codebases intelligently

For detailed language support and exclusion rules, see the full documentation.

---

🎯 Common Commands

# Default: creates .securevibes/scan_report.md (markdown format)
securevibes scan .

# Export JSON for CI/CD pipeline
securevibes scan . --format json --output security-report.json

# Custom markdown report (saved to .securevibes/custom_report.md)
securevibes scan . --format markdown --output custom_report.md

# Terminal table output (no file saved)
securevibes scan . --format table

# Focus on critical/high severity
securevibes scan . --severity high

# Fast scan with cheaper model
securevibes scan . --model haiku

# Quiet mode for automation
securevibes scan . --quiet

# Run individual sub-agents
securevibes scan . --subagent assessment
securevibes scan . --subagent code-review
securevibes scan . --subagent report-generator

# DAST (optional): skill‑gated dynamic validation
securevibes scan . --subagent dast --target-url http://localhost:3000
  # Validates only when a matching skill is available (e.g., IDOR)
  # Writes .securevibes/DAST_VALIDATION.json; no ad‑hoc files in repo

# Agentic detection override
securevibes scan . --agentic      # Force ASI threat requirement
securevibes scan . --no-agentic   # ASI threats optional

# Works with any supported language:
securevibes scan /path/to/go-app        # Go project
securevibes scan /path/to/ruby-app      # Ruby project
securevibes scan /path/to/mixed-stack   # Multi-language project

# PR review (requires prior scan artifacts in .securevibes/)
securevibes pr-review . --base main --head feature-branch
securevibes pr-review . --range abc123~1..abc123
securevibes pr-review . --diff changes.patch

# Commit tracking (requires baseline scan and scan_state.json)
securevibes pr-review . --since-last-scan
securevibes pr-review . --since 2026-02-01
securevibes pr-review . --last 10

# Update base artifacts from PR findings
securevibes pr-review . --range abc123~1..abc123 --update-artifacts

# Clean transient PR artifacts before reruns
securevibes pr-review . --range abc123~1..abc123 --clean-pr-artifacts

# PR review output formats (default: markdown)
securevibes pr-review . --base main --head feature-branch --format markdown
securevibes pr-review . --base main --head feature-branch --format json --output pr_review.json
securevibes pr-review . --base main --head feature-branch --format table

# PR review severity threshold
securevibes pr-review . --base main --head feature-branch --severity high

# Catchup: pull latest + review since last full scan
securevibes catchup . --branch main

securevibes catchup requires a clean working tree (commit, stash, or discard local changes first).

PR review runtime controls:

# Timeout per PR review attempt in seconds (default: 240)
export SECUREVIBES_PR_REVIEW_TIMEOUT_SECONDS=300

# Number of PR review attempts before giving up (default: 4)
export SECUREVIBES_PR_REVIEW_ATTEMPTS=5

PR review fails closed if diff context would be truncated (more than 16 prioritized files or any hunk over 200 lines). PR review also fails closed when PR_VULNERABILITIES.json is not produced after retry attempts. Split large reviews with smaller --range, --last, or --since windows.

PR review artifacts (written to .securevibes/):

• DIFF_CONTEXT.json (parsed diff summary)
• PR_VULNERABILITIES.json (raw findings)
• pr_review_report.md (default markdown report)
• scan_state.json (commit tracking for pr-review/catchup)

---

⚙️ Configuration

Models

SecureVibes uses a three-tier priority system for model selection:

Priority Hierarchy:

1. 🥇 Per-agent environment variables (highest)
2. 🥈 CLI --model flag (applies to all agents)
3. 🥉 Default "sonnet" (fallback)

Examples:

# All agents use haiku
securevibes scan . --model haiku

# All use haiku, except code-review uses opus
export SECUREVIBES_CODE_REVIEW_MODEL=opus
securevibes scan . --model haiku

# Fine-grained control per agent
export SECUREVIBES_ASSESSMENT_MODEL=haiku
export SECUREVIBES_CODE_REVIEW_MODEL=opus
securevibes scan .  # Others use default (sonnet)

Models: haiku (fast/cheap), sonnet (balanced), opus (thorough/expensive)

Per‑Agent Overrides

Override specific agent models via environment variables:

# Authenticate first (see Quick Start above)

# Override specific agent models (overrides CLI --model flag)
export SECUREVIBES_CODE_REVIEW_MODEL="opus"  # Max accuracy
export SECUREVIBES_THREAT_MODELING_MODEL="sonnet"
export SECUREVIBES_PR_CODE_REVIEW_MODEL="sonnet"

# Control analysis depth (default: 50)
export SECUREVIBES_MAX_TURNS=75  # Deeper analysis

---

🐍 Python API (minimal)

import asyncio
from securevibes import Scanner

async def main():
    # Auth via Claude Agent SDK:
    # - ANTHROPIC_API_KEY or
    # - Session token from `claude` CLI (/login)
    scanner = Scanner(
        model="sonnet",  # Use shorthand: sonnet, haiku, opus
        debug=True  # Show agent narration for verbose output
    )
    
    result = await scanner.scan("/path/to/repo")
    print(f"Found {len(result.issues)} vulnerabilities")
    print(f"Cost: ${result.total_cost_usd:.4f}")

asyncio.run(main())

---

📚 Full Documentation

This is a quick reference for PyPI users. For comprehensive documentation, visit:

🌐 securevibes.ai — Full website with docs & blog

Including:

• 📖 How It Works
• ✨ Features
• 📝 Blog

---

👤 Author

Built by @anshumanbh

🌐 securevibes.ai — Website, docs & blog

🌟 Star the repo to follow development!

---

🙏 Acknowledgments

• Powered by Claude by Anthropic
• Built with Claude Agent SDK
• Inspired by traditional SAST tools, reimagined with AI

---

License: AGPL-3.0 | Requires: Python 3.10+