Connect Firebreak to Claude, Cursor, or any MCP-compatible client. Then just ask:
You: "Is my app secure?"
Claude: [calls firebreak tools] Found 6 vulnerabilities. 1 high (CORS misconfiguration),
5 medium (missing security headers). Security score: B. Want me to fix them?
No CLI to learn. No reports to read. The AI runs the scans, interprets the results, and walks you through the fixes.
Security MCP server that turns your AI into a penetration tester. 20 tools, 100 best practice guides, 47 vulnerability patterns. Connect to Claude, Cursor, or any MCP client.
星數
4
工具
—
可用率
—
Why Firebreak
AI-generated code ships with predictable vulnerabilities: inconsistent auth middleware, permissive RLS policies, IDOR on every endpoint, secrets in the JS bundle. Developers who use AI to build don't always know how to test what it builds.
Firebreak flips the problem. Instead of expecting developers to learn security tooling, it gives the AI the security tooling. The same AI that wrote the code can now test it, find the holes, and fix them.
What makes it different:
MCP-native — not a CLI wrapper. Built from the ground up for AI tool calling.
VCVD — 47 vulnerability patterns specific to AI-generated code that traditional scanners miss.
Proof over theory — every finding includes the actual HTTP request/response that proves the vulnerability.
Safe by design — rate limited, scope-locked, non-destructive. Can't accidentally DROP your database.
Quick Start
Option 1: Build from source
git clone https://github.com/protonese3/Firebreak.git
cd Firebreak
cargo build --release
./target/release/firebreak
Option 2: Docker
git clone https://github.com/protonese3/Firebreak.git
cd Firebreak
docker compose up -d
Restart Claude Desktop. You now have 20 security tools available.
Connect to Cursor / Windsurf / Any MCP Client
Point your client's MCP configuration to http://localhost:9090/mcp. The server speaks standard MCP (JSON-RPC 2.0 over HTTP).
Verify it works
Ask your AI: "What security tools do you have available?"
It should list the Firebreak tools. Then try: "Scan https://httpbin.org for security issues"
How It Works
The AI is the orchestrator. Firebreak doesn't decide what to test — it provides the tools. The AI picks the strategy based on what the user asks and what it finds along the way.
Typical flow
1. User: "Test my app at https://myapp.com"
2. AI calls firebreak_scan_quick({ target_url: "https://myapp.com" })
└── Firebreak probes headers, paths, CORS, TLS
└── Returns: 4 findings, score B
3. AI explains results in plain language
└── "Found a CORS misconfiguration and 3 missing headers..."
4. User: "Fix the CORS issue"
5. AI calls firebreak_finding_fix({ finding_id: "...", framework: "express" })
└── Returns: before/after code diff for Express
6. User applies fix
7. AI calls firebreak_replay({ finding_id: "..." })
└── Returns: "Fixed. Server no longer reflects arbitrary origins."
Tool Reference
Knowledge Tools
These don't hit any external service. They query Firebreak's built-in security knowledge base.
Tool
Description
Example
firebreak_best_practice
Security best practices for a topic
{ topic: "jwt-auth" }
firebreak_check_pattern
Scan code for insecure patterns
{ code: "...", language: "javascript" }
firebreak_explain_vuln
Explain a vulnerability type
{ vuln_id: "IDOR" }
firebreak_security_checklist
Generate a security checklist
{ stack: ["nextjs", "supabase"] }
firebreak_owasp_check
Map a finding to OWASP Top 10
{ description: "SQL injection in login" }
firebreak_analyze_rls
Analyze SQL for RLS policy issues
{ sql: "CREATE POLICY..." }
Scan Tools
These make HTTP requests to the target. Rate limited and scope-locked.
Stack traces, debug info, internal paths in error responses
Subdomain reference discovery
Subdomains referenced in HTML, JS, and headers
Cache header security
Cache-Control, Pragma, Expires on sensitive responses
VCVD — Vibe Coding Vulnerability Database
47 vulnerability patterns that AI-generated code gets wrong. Traditional scanners don't look for these because they're specific to how LLMs write code.
Auth & Identity
ID
Pattern
Severity
VC-AUTH-001
Inconsistent auth middleware — AI applies auth to some routes but not others
CRITICAL
VC-AUTH-002
Client-only validation — role checks exist in React but not in the API
CRITICAL
VC-AUTH-003
JWT decoded without verification — jwt.decode() instead of jwt.verify()
CRITICAL
VC-AUTH-004
Service key in client code — Supabase service_role key in the JS bundle
CRITICAL
VC-AUTH-005
Token expiry >30 days — "for convenience"
HIGH
VC-AUTH-006
OAuth without state parameter
HIGH
VC-AUTH-007
Password in URL query string
HIGH
VC-AUTH-008
Session not regenerated after login
MEDIUM
Data Access
ID
Pattern
Severity
VC-DATA-001
IDOR — sequential IDs without ownership check
CRITICAL
VC-DATA-002
USING (true) in RLS — AI's "temporary" policy that ships to prod
CRITICAL
VC-DATA-003
Table created without enabling RLS
CRITICAL
VC-DATA-004
SELECT * exposing password hashes and PII
HIGH
VC-DATA-005
...req.body spread into DB insert without field whitelist
HIGH
VC-DATA-006
GraphQL introspection enabled in production
MEDIUM
VC-DATA-007
Unbounded N+1 queries as DoS vector
MEDIUM
VC-DATA-008
Multi-tenant queries missing tenant_id filter
CRITICAL
Injection
ID
Pattern
Severity
VC-INJ-001
Template literals in SQL queries
CRITICAL
VC-INJ-002
innerHTML with user input
HIGH
VC-INJ-003
User content stored and rendered without sanitization
CRITICAL
VC-INJ-004
User input in exec() / spawn()
CRITICAL
VC-INJ-005
File paths built from user input
HIGH
VC-INJ-006
Server-side fetch with user-provided URL (SSRF)
HIGH
VC-INJ-007
User input passed as template source
HIGH
VC-INJ-008
MongoDB operators in JSON input
HIGH
Infrastructure
ID
Pattern
Severity
VC-INFRA-001
Debug mode / stack traces in production
HIGH
VC-INFRA-002
Access-Control-Allow-Origin: * with credentials
HIGH
VC-INFRA-003
Missing HSTS, X-Frame-Options, CSP
MEDIUM
VC-INFRA-004
Unnecessary Docker ports exposed
MEDIUM
VC-INFRA-005
User uploads stored without encryption
MEDIUM
VC-INFRA-006
TLS 1.0/1.1 still enabled
MEDIUM
VC-INFRA-007
Admin panel accessible from public internet
HIGH
VC-INFRA-008
Default credentials unchanged
CRITICAL
Frontend
ID
Pattern
Severity
VC-FE-001
API keys in JavaScript bundle
CRITICAL
VC-FE-002
Auth guard only in React Router, not in API
HIGH
VC-FE-003
JWT stored in localStorage
MEDIUM
VC-FE-004
Forms without CSRF protection
HIGH
VC-FE-005
Redirect URL from query param without validation
MEDIUM
VC-FE-006
postMessage handler without origin check
MEDIUM
VC-FE-007
Source maps accessible in production
LOW
VC-FE-008
Sensitive fields without autocomplete="off"
LOW
Security Scoring
Every scan produces a letter grade based on what was found:
Grade
Criteria
Interpretation
A
Zero critical or high. Max 2 medium.
Ship it.
B
Zero critical. Some high or medium.
Fix the highs soon.
C
Multiple high severity issues.
Needs work before production.
D
1-2 critical or >5 high.
Significant risk. Prioritize fixes.
F
3+ critical or a full compromise chain.
Do not deploy.
Safety Guardrails
Firebreak is designed for authorized testing only.
Guardrail
How it works
Rate limiting
10 requests/second to the target (configurable). Sliding window.
Scope lock
Only attacks the specified target URL. No lateral movement, no subdomain enumeration.
Non-destructive
No DELETE, DROP, UPDATE, or POST requests that modify data. Probing only.
Consent
The AI asks for confirmation before the first scan.
Audit trail
Every HTTP request is logged with timestamp, target, and reason.
The release binary is ~8 MB with no runtime dependencies.
Contributing
Contributions are welcome. Here's how to get involved:
Getting started
git clone https://github.com/protonese3/Firebreak.git
cd Firebreak
cargo build
cargo test
The server runs on http://localhost:9090/mcp by default.
Areas where help is needed
New VCVD patterns — Found a vulnerability pattern that AI consistently generates? Add it to src/vcvd/data.rs. Each pattern needs an ID, description, severity, detection hint, and fix.
Scan engine checks — Add new vulnerability checks in src/engine/checks.rs. Each check function takes a reqwest client, target URL, and safety reference. Return a Vec<Finding>.
Best practice guides — Add markdown files to knowledge/best-practices/. Then register them in src/tools/knowledge/best_practices.rs.
Framework-specific fixes — The finding_fix tool generates fix code. Add framework support in src/tools/analysis.rs in the framework_fix_example function.
Dashboard — The React frontend lives in frontend/. Run npm run dev for the dev server.
Submitting changes
Fork the repo
Create a branch (git checkout -b add-new-check)
Make your changes
Run cargo check && cargo clippy -- -D warnings
Open a pull request with a clear description of what you changed and why
Code style
No unnecessary comments. If the code says what it does, don't add a comment.
Guard clauses over nested if/else.
Match the style of existing code.
Every finding must have verifiable evidence — no "this might be vulnerable."
Reporting security issues
If you find a security vulnerability in Firebreak itself, please email [email protected] instead of opening a public issue. We'll respond within 48 hours.
Do I need to be a security expert to use this?
No. That's the point. The AI handles the security expertise. You just tell it what to test.
What MCP clients work with Firebreak?
Any client that supports MCP over HTTP: Claude Desktop, Claude Code, Cursor, Windsurf, and others.
Is it safe to run against production?
Firebreak is non-destructive (read-only probing), but you should always test against staging first. Rate limiting is on by default.
Can I add my own vulnerability checks?
Yes. Add check functions in src/engine/checks.rs and wire them into the scan methods in src/engine/mod.rs.
Why Rust?
Single binary with no runtime dependencies. Fast. Memory safe. We ship one file and it works.
Why AGPL-3.0?
If you run Firebreak as a service for others, you must share your modifications. If you use it internally or self-host for your own team, you don't need to do anything special.
If you build a commercial service on top of Firebreak, the AGPL requires you to open-source your modifications. For internal and self-hosted use, no restrictions beyond the standard AGPL terms.
F
Firebreak
synthet1cc·4 顆星·
Security MCP server that turns your AI into a penetration tester. 20 tools, 100 best practice guides, 47 vulnerability patterns. Connect to Claude, Cursor, or any MCP client.
上次檢查時間:2026-10-27。 Security MCP server that turns your AI into a penetration tester. 20 tools, 100 best practice guides, 47 vulnerability patterns. Connect to Claude, Cursor, or any MCP client.