chiprook
← Security
SecurityOctober 24, 2026, 11:00

MCP tool poisoning: 36.5% attack success across 20 models

Hidden instructions in MCP tool descriptions bypass code review: the MCPTox benchmark found poisoned tools succeeded 36.5% of the time on average across 20 models, peaking at 72.8% against one. Nine of 11 public MCP registries accepted poisoned tools with no automated screening. Defences include the free MCP-Scan scanner, version pinning and least privilege.

MCP tool poisoning: 36.5% attack success across 20 models
#MCP#InvariantLabs#Cursor#MCP-Scan
Read next
Security

MCP tool poisoning: hidden instructions in tool descriptions hijack AI agents

Security

Tool Poisoning on MCP Servers: The Attack Vector Nobody's Patching

Security

Memory poisoning in AI agents persists across sessions and fires weeks later

Security

MCP becomes the top attack surface in developer tooling