MCP Security at Wiz with Rami McCarthy

What It Covers

Rami McCarthy from Wiz discusses Model Context Protocol security risks, AI-generated code vulnerabilities, secrets leakage patterns, GitHub Actions supply chain attacks, and practical security guidance for organizations adopting AI development tools.

Key Questions Answered

• •AI Secrets Leakage: Four out of five new public repository secrets are AI-related, driven by LLMs generating hardcoded credentials and AI tools using plain text configuration files without established secret scanning coverage for emerging platforms.
• •MCP Local Server Risk: Organizations should maintain small allowlists of approved MCP servers tied to reputable organizations, treating them like Chrome extensions with permissioning models rather than allowing unrestricted consumption of 4,000+ available servers from unvetted sources.
• •Auto-Approval Vulnerability: Running MCP clients with auto-approval enabled eliminates human oversight that catches indirect prompt injection attacks, where malicious prompts embedded in GitHub issues can trigger unintended actions without user intervention or detection.
• •Supply Chain Attack Pattern: The TJ Actions compromise demonstrates multi-step attacks where attackers compromise less popular actions to gain access to widely-used dependencies, requiring organizations to monitor not just direct dependencies but entire upstream chains.