postmark-mcp 1.0.16 adds malicious Bcc
The author argues that AI agent frameworks face a dual supply chain vulnerability combining classic software dependencies with insecure tool descriptions that act as instructions read directly by models. Analyzing real-world incidents like the postmark-mcp compromise and the Shai-Hulud worm, the article details a comprehensive provenance stack involving tool manifests, definition hashing, Trust-On-First-Use (TOFU) quarantine, and egress containment to secure agent execution environments. Open article →