Findings & checks · 1 flagged
MEDIUM"Never tell the user" non-disclosure imperative in the skillSS-SKILL-INJECT-IMPERATIVE-01 · Prompt injection · SECURITY.md
A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.
- Delete any "never reveal / never tell the user" line aimed at the model's own behavior.
- Keep legitimate confidentiality rules about external data (e.g. "do not echo API keys"), which are user-protective, not user-deceiving.
Trace & refs
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
AI Security Crew
A lightweight MCP server for security reviews built for vibe coding — injects security requirements prior to code generation, scans dependencies for CVEs, and verifies generated code, all without breaking your coding rhythm.
Jump to installation:
- MCP Server — full feature set with Jira, Confluence, CVE scanning (with reachability), and threat modeling
- Claude Code Plugin — install 3 security skills globally in Claude Code (no Jira/MCP needed)
- Claude Code Skills only — manually add slash commands to a specific project
Claude Code Plugin
Install all three security skills directly into Claude Code — no MCP server, no Jira, no configuration required.
/plugin install Srajangpt1/ai_security_crewThis gives you three commands available in any project:
| Command | When to use |
|---|---|
/sec-review | Before coding — get risk level, OWASP guidelines, and a security prompt for AI code generation |
/verify-code | After coding — review code for vulnerabilities with a checklist and prioritized fixes |
/threat-model | For new features — identify threats with evidence links, mitigations, and optional threat-model.md |
Claude Code Skills
If you prefer to add the skills to a specific project only (instead of globally), clone this repo and the slash commands in .claude/commands/ are available automatically in Claude Code when working in the project directory.
Tools
Pre-coding
| Tool | When to Use |
|---|---|
lightweight_security_review | Before any coding task — get security requirements and guidelines for your tech stack |
assess_ticket_security | Before coding from a Jira ticket — pull security requirements directly from the ticket |
perform_threat_model | For significant new features — generate a structured threat model (STRIDE, attack surfaces) |
Dependency security
| Tool | When to Use |
|---|---|
verify_packages | When adding packages — confirm they exist with valid versions (catches hallucinated package names) |
scan_dependencies | When adding packages — scan for CVEs and check reachability in your code |
Post-coding
| Tool | When to Use |
|---|---|
verify_code_security | After generating code — AI-powered security review against OWASP guidelines |
Threat model persistence
| Tool | When to Use |
|---|---|
search_previous_threat_models | Before creating a new threat model — check if one already exists in Confluence |
update_threat_model_file | After perform_threat_model — write the threat model to threat-model.md in the repo |
Agent Workflow
The server automatically sends workflow instructions to any connecting agent (Claude, Cursor, etc.) via the MCP initialize handshake. Agents will follow this workflow without additional configuration:
- Before coding — call
lightweight_security_review(orassess_ticket_securityfor Jira tickets) - When adding packages — call
verify_packages, thenscan_dependencieswith the code that uses them - After generating code — call
verify_code_securityand follow thereview_promptto report findings - For significant features — call
perform_threat_modeland persist withupdate_threat_model_file
Dependency Scanning
scan_dependencies uses OSV.dev to find CVEs and performs reachability analysis to determine if vulnerable code paths are actually called:
| Status | Meaning |
|---|---|
reachable | Vulnerable function is called in your code — action required |
not_reachable | Vulnerable function is not called |
not_imported | Package is not imported at all |
uncertain | AI analyzed the code but could not determine reachability |
no_code_provided | No code snippets were passed to the tool |
Reachability is determined by (in order): OSV function-level symbols → keyword matching against the vuln summary → AI analysis via ctx.sample().
Quick Start
1. Build the image
docker build -t mcp-security-review:latest .2. Configure your IDE
Add to your MCP config (Claude Desktop, Cursor, etc.):
{
"mcpServers": {
"sec-review": {
"command": "docker",
"args": [
"run", "--rm", "-i",
"-e", "JIRA_URL",
"-e", "JIRA_USERNAME",
"-e", "JIRA_API_TOKEN",
"-e", "CONFLUENCE_URL",
"-e", "CONFLUENCE_USERNAME",
"-e", "CONFLUENCE_API_TOKEN",
"mcp-security-review:latest"
],
"env": {
"PATH": "/usr/local/bin:/usr/bin:/bin",
"JIRA_URL": "https://your-domain.atlassian.net",
"JIRA_USERNAME": "[email protected]",
"JIRA_API_TOKEN": "your-token"
}
}
}
}Authentication
Supported methods:
- API Token (Jira/Confluence Cloud):
JIRA_API_TOKEN,CONFLUENCE_API_TOKEN - Personal Access Token (Server/Data Center):
JIRA_PERSONAL_TOKEN,CONFLUENCE_PERSONAL_TOKEN - OAuth 2.0 (Cloud): run
docker run --rm -it mcp-security-review:latest --oauth-setup
HTTP Transport
Run as a persistent HTTP service instead of stdio:
# Streamable HTTP (recommended)
docker run --rm -p 8000:8000 mcp-security-review:latest --transport streamable-http
# SSE
docker run --rm -p 8000:8000 mcp-security-review:latest --transport sseSecurity Guidelines
Includes 101 OWASP Cheat Sheets loaded automatically into security assessments. Add your own org-specific guidelines:
python3 scripts/add_custom_guideline.pyOr manually create markdown files in src/mcp_security_review/security/guidelines/docs/:
category: your_category
priority: high
tags: tag1, tag2, tag3
# Your Guideline Title
...See docs/ADDING_CUSTOM_GUIDELINES.md for details.
Contributing
- Check CONTRIBUTING.md for development setup.
- Make changes and submit a pull request.
Pre-commit hooks enforce code quality (Ruff, Prettier, Pyright). Run uv run pytest before submitting.
Security
Never commit API tokens. See SECURITY.md for best practices.
License
Licensed under MIT — see LICENSE.