CrowdStrike Publishes Taxonomy of Agentic Tool Chain Attacks: Tool Poisoning, Shadowing, and Rugpull
CrowdStrike·high signal
CrowdStrike published a technical breakdown naming three distinct agentic attack classes: tool poisoning (injecting malicious instructions into tool descriptions), tool shadowing (overriding legitimate tools with malicious lookalikes), and rugpull attacks (tools that behave legitimately until a trigger condition). The post frames MCP server compromise as an architectural supply chain risk — every agent trusting a compromised server inherits the attack. Recommended defenses include signed manifests, version pinning, and explicit upgrade approval.