Research

What Is an MCP Rug Pull? How Poisoned AI Tools Hijack Agents | Logic42 Practice Brief

Page published

Earliest dated coverage: 18 Oct 2026 · First observed: 17 Nov 2026 · Latest dated coverage: 18 Oct 2026

Coverage timeline

18 Oct 2026logic42.ai

Single-source research — one report is available.

Why it matters

MCP tool descriptions and A2A Agent Cards can become prompt-injection channels after initial approval, making metadata integrity checks and limits on agent access to secrets important defensive controls.

Logic42 reports testing 1,000 simulated MCP/A2A supply-chain attacks involving mutable tool descriptions, poisoned Agent Cards, and cross-server tool shadowing. In one reported sandbox demonstration, a changed tool description induced a coding agent to read an SSH private key and transmit it through a telemetry query parameter. The brief claims cryptographic integrity checks and isolation blocked the tested attacks, but the provided excerpt does not independently substantiate its performance figures.