New Technique Medium TD-2026-0040 For CISO, AppSec, AI Eng, Red Team, SOC / IR AI-assisted
MCP tool poisoning hits 22.9% of tested agent cells across 12 frameworks
An arXiv measurement study tested MCP tool poisoning with real tool execution across 12 agentic AI frameworks, 18 models, and 45 attack scenarios. The overall attack success rate was 22.9% (2,113 vulnerable cells). Of those vulnerable cells, 12.1% verifiably exfiltrated a secret.
Why it matters
Any organization running agents that connect to MCP tool servers can have model context altered by tool descriptions, schema fields, or response data. An attacker who controls that content can trigger unintended tool calls and steal secrets, and the study found no framework with intentional defense.
Do this
- Inventory all MCP servers and agent frameworks in use, and record which ones pass tool descriptions and responses into model context.
- Restrict outbound network egress from agent runtimes so that unapproved external endpoints cannot receive data.
- Require human approval or policy checks before agent tool calls that touch secrets, files, or external network destinations.
- Do not rely on regex metadata scanning alone; the study reports it rejected 20 of 45 scenarios with a 7.2% false-positive rate.
- Test your own agent stack in a non-production environment, for example with the open-source mcpvenom tool, before deploying new MCP servers.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“Of these, 64.9% dispatch an attacker tool call, 23.0% contact the attacker endpoint, and 12.1% verifiably exfiltrate a secret.”
“Resistance is model-level: nine core frameworks cluster within 23 to 26% (chi-squared = 3.30, p = 0.91), and no framework implements intentional defense.”
“a regex description scanner rejects 20 of 45 scenarios at a 7.2% false-positive rate, so metadata scanning alone is a partial control.”
Sources
Affected / related
MCP tool poisoning, Model Context Protocol (MCP), mcpvenom
Published 2026-11-02 18:17 UTC · updated 2026-11-08 17:53 UTC · AI-assisted summary, checked against quoted sources · How we score