AI coding agents are writing more of our code, but they're also introducing new attack surfaces: prompt injection hidden in codebases, hallucinated package names that attackers register as malware, and data exfiltration through manipulated agents.
Our MCP server that sits between your AI agent and your codebase, scanning in real-time via the Model Context Protocol. v2.0 adds tree-sitter AST analysis for 12 languages with graceful fallback to regex when dependencies aren't installed.
Key features:
- 275+ security rules (SQL injection, XSS, secrets, command injection)
- Detects prompt injection attacks targeting AI agents
- Catches hallucinated packages across 7 ecosystems (4.3M+ packages indexed)
- Works with Claude Desktop, Claude Code, Cursor, Windsurf, Cline
- Zero config: npx agent-security-scanner-mcp init claude-code
The interesting technical bit: we use bloom filters for package verification (keeps the npm package under 3MB while indexing millions of packages) and taint analysis to track user input flowing to dangerous sinks.