ARC Security MCP Server
Security intelligence built for AIs, not humans. Any agent can connect and query: "Is this skill safe?"
Built from real audit data — 908+ ClawHub skills surveyed across 87 audit rounds, 5,411+ findings (806 critical, 1,377 high), 25 attack classes mapped to OWASP Agentic AI Top 10. Five active responsible disclosures with framework maintainers.
Install
Quickstart (no install needed):
npx arc-security-mcp
Claude Desktop / Claude Code (remote SSE):
{
"mcpServers": {
"arc-security": {
"url": "https://arcself.com/mcp/sse"
}
}
}
Claude Desktop / Claude Code (local via npm):
npm install -g arc-security-mcp
{
"mcpServers": {
"arc-security": {
"command": "arc-security-mcp",
"args": []
}
}
}
Security Scanning Tools
Check if a ClawHub skill is safe to install. Checks our curated database first, then auto-fetches and scans unknown skills in real-time.
Fetch any skill from ClawHub and run a full security scan: 31 static patterns + AI-powered intent analysis.
Static analysis against 31 pattern rules covering shell injection, credential exposure, identity manipulation, supply chain risks, and APT-grade evasion techniques.
AI reads the SKILL.md and detects capability-purpose mismatches, exfiltration channels, memory poisoning, and distributed attack chains.
Detailed information on any of 25 documented attack classes, with OWASP Agentic AI Top 10 mapping.
Full mapping between our 25 attack classes and OWASP Agentic AI Top 10 (ASI01-ASI10).
All known dangerous code patterns with descriptions, real-world examples, and mitigations.
Current AI agent security threat landscape summary with ecosystem statistics.
Tailored security checklist for a specific skill type (financial, communication, filesystem, database, browser, shell).
Runtime Monitoring v0.5
The first EDR (Endpoint Detection and Response) built for AI agents. Track tool calls, file access, and network activity in real-time.
Start monitoring an agent session. Returns a session token for tracking.
Report a tool call, file access, network request, or shell command for real-time risk assessment.
End monitoring and get a full session security report with severity breakdown.
Define behavioral rules for a skill: allowed/denied tools, file patterns, network targets, rate limits.
Query all security alerts for a running monitoring session.
Knowledge Base
- 87 ClawHub audit rounds — 908+ skills surveyed, 433+ deep-scanned
- 5,411+ findings (806 critical, 1,377 high) in curated database
- 115 attack patterns across 25 attack classes, mapped to OWASP Agentic AI Top 10
- 7 runtime detection rules — exfiltration, sensitive file access, dangerous shell, capability escalation
- LLM-powered intent analysis — detects semantic threats no regex can catch
- 5 active responsible disclosures with Anthropic, Microsoft, FalkorDB, deepset, and others