██████╗ ██████╗ ███████╗███╗   ██╗███████╗███████╗ ██████╗ 
██╔═══██╗██╔══██╗██╔════╝████╗  ██║██╔════╝██╔════╝██╔════╝ 
██║   ██║██████╔╝█████╗  ██╔██╗ ██║███████╗█████╗  ██║      
██║   ██║██╔═══╝ ██╔══╝  ██║╚██╗██║╚════██║██╔══╝  ██║      
╚██████╔╝██║     ███████╗██║ ╚████║███████║███████╗╚██████╗ 
 ╚═════╝ ╚═╝     ╚══════╝╚═╝  ╚═══╝╚══════╝╚══════╝ ╚═════╝ 

Find vulnerabilities. Prove they're real. Patch them. Prove the patch works.
Free forever. No API keys. Runs on your machine.

Install · The prove loop · How it works · CLI reference · Models

---

The problem with every other security tool

They give you a list of 47 findings.
You fix maybe 3 of them — the ones you're sure about.
The other 44 sit there because you can't tell what's real.

OpenSec Intelligence solves the trust problem.

It doesn't just find the vulnerability. It runs an exploit to prove it exists, patches it with a local AI, then runs the same exploit again to prove the patch closed the hole. You see exactly what happened at every step.

npm install -g opensec-intelligence
ollama pull qwen2.5-coder:14b
opensec scan ./

---

The prove loop

This is what makes OpenSec different from every other free tool.

opensec prove ./

  ╭─────────────────────────────────────────────────────────────╮
  │  PROVING: SQL Injection in src/db/queries.js:47             │
  │                                                             │
  │  🔴 VULNERABILITY CONFIRMED                                 │
  │     Input:  "' OR '1'='1"                                  │
  │     Result: returned 3 rows — authentication bypassed       │
  │                                                             │
  │  ─────────────────────────────────────────────────────────  │
  │                                                             │
  │  ✅ PATCH VERIFIED — exploit no longer works                │
  │     Same input: "' OR '1'='1"                              │
  │     Result:     parameterized query — 0 rows                │
  ╰─────────────────────────────────────────────────────────────╯

What's happening:

1. A local LLM writes a minimal exploit script for the finding
2. The script runs in a sandboxed subprocess and confirms the vuln fires (🔴)
3. The Fixer agent writes and applies the patch
4. The same exploit runs again — it must fail (✅)
5. Proof saved to .opensec/proofs/ for audit

No other free, local tool does this. Most enterprise tools don't either.

---

Quickstart

# Install
npm install -g opensec-intelligence

# Pull a model (free, runs locally — one time)
ollama pull qwen2.5-coder:14b

# Scan your repo
opensec scan ./

# See it on a deliberately vulnerable demo app
opensec scan --demo

No account. No API key. Nothing leaves your machine.

---

How it works

Your codebase
      │
      ▼
┌─────────────┐
│   Scanner   │  Walks every file — code, Dockerfiles, k8s, .env,
│             │  Terraform, CI/CD — in under 30 seconds.
│             │  25+ deterministic rules. No LLM needed here.
└──────┬──────┘
       │ structured findings JSON
       ▼
┌─────────────┐
│   Analyst   │  Sends each HIGH/CRITICAL finding to a local LLM
│             │  with ±15 lines of file context.
│             │  Filters anything below 0.7 confidence.
└──────┬──────┘
       │ confirmed findings
       ▼
┌─────────────┐
│  Consensus  │  Adversarial second pass on every CRITICAL.
│             │  Tries to argue the finding ISN'T a vulnerability.
│             │  If it still holds up — it's real.
└──────┬──────┘
       │ validated findings
       ▼
┌─────────────┐
│    Fixer    │  Writes the exact patch.
│             │  Shows you a colored before/after diff.
│             │  [A]pply / [S]kip / [Q]uit — you decide.
│             │  Git commits on approval.
└─────────────┘

The insight that makes this work: single-model tools hallucinate. If one model says something is a vulnerability, it might be wrong. OpenSec's consensus layer requires independent confirmation — if models disagree, you don't get paged.

---

What gets scanned

Category	File types
Code	.py .js .ts .go .rb .java .php .rs .cpp .cs
Infrastructure	Dockerfile docker-compose.yml *.tf *.hcl *.toml
Kubernetes / Config	*.yaml *.yml openapi.* *.json
Secrets	.env .env.* *.pem *.key *.p12 *.pfx
CI/CD	.github/workflows/*.yml .gitlab-ci.yml Jenkinsfile

What gets detected: AWS/GitHub/Stripe keys, hardcoded passwords, SQL injection, command injection, path traversal, CORS wildcards, weak crypto (MD5/SHA1), eval usage, Docker root users, k8s privileged containers, Terraform open security groups, and more.

---

Commands

# Scan
opensec scan ./              # Full 4-agent scan
opensec scan ./ --quick      # Deterministic scanner only (no LLM, instant)
opensec scan ./ --ci         # JSON output, exit 1 on CRITICAL — use in CI/CD
opensec scan --demo          # Run on bundled vulnerable app — see the pipeline live

# Prove + fix
opensec prove ./             # Generate exploit → confirm → patch → re-confirm
opensec prove ./ --dry-run   # Generate and run exploit, skip patching
opensec prove ./ --show-exploit  # Print the LLM-generated exploit code
opensec fix                  # Apply patches from last scan interactively

# Report + serve
opensec report               # HTML security report
opensec serve --port 4141    # Start HTTP + MCP server

# Model
opensec -m llama3.2:3b       # Override Ollama model

---

CI/CD integration

# .github/workflows/security.yml
opensec scan ./ --ci --output json
# exits 1 if any CRITICAL findings — blocks the merge

See .github/workflows/opensec.yml for a full example.

---

MCP integration

OpenSec exposes an MCP server so any MCP-compatible tool (Claude Code, goose, Cursor) can call it:

opensec serve --port 4141
# connect any MCP client to http://localhost:4141/v1/stream

Tool	What it does
scan_repo	Trigger a full or quick scan
get_findings	Structured JSON of last scan results
apply_fix	Apply a specific validated fix
prove_finding	Run the prove loop on a specific finding
get_report	Generate HTML report

---

Model recommendations

Model	RAM	Best for
qwen2.5-coder:14b	16GB	Best overall — recommended
deepseek-r1:14b	16GB	Best for the consensus reasoning pass
codellama:13b	16GB	Balanced speed + accuracy
llama3.2:3b	8GB	Fast scanner on low RAM machines

ollama pull qwen2.5-coder:14b   # recommended
ollama pull llama3.2:3b          # minimum spec

---

Local vs Cloud

Mode	Cost	Privacy	Command
Local (default)	Free	100% private	opensec scan ./
Quick	Free	100% private	opensec scan ./ --quick
Cloud	API cost	Analyst + Consensus only	opensec scan ./ --cloud

Cloud mode keeps Scanner and Fixer fully local. Only the confirmation passes use cloud models.

---

Safety

The prove loop runs exploit code in a sandboxed subprocess:

• Real credentials and environment variables are never passed to the subprocess
• eval() and new Function() string codegen are disabled inside the subprocess
• Network access is blocked before execution
• Filesystem writes outside /tmp are blocked
• Hard 10-second timeout
• Exploit files are deleted immediately after use

⚠️ Do not run opensec prove on untrusted code. File content is embedded in the LLM prompt — a crafted file could influence the generated exploit. Full isolation requires a container.

---

What's coming

• VS Code extension — inline vulnerability highlights as you type
• GitHub App — auto-scan every PR, post findings as review comments
• SARIF output — GitHub Advanced Security integration
• Multi-model parallel scanning (--multi-model)
• Web dashboard for team findings
• Slack / Discord alerts

Pro + Enterprise tiers are in development for teams that need zero-false-positive guarantees, air-gapped deployment, and compliance reports (SOC2, ISO 27001, OWASP). Join the waitlist →

---

Contributing

Security rules live in src/security/patterns.ts — adding a new detection pattern is a 5-line PR. See CONTRIBUTING.md.

git clone https://github.com/prabindersinghh/opensec-intelligence
npm install
npm test          # 148 tests
opensec scan ./   # dogfood

---

MIT License · npm · Issues

Built by Prabinder Singh — B.Tech CS, Thapar Institute · Founder, Leorit.ai