New Technique Medium TD-2026-0045 For AppSec, AI Eng, CISO, SOC / IR AI-assisted
MCP destructive tools often lack human approval on server and host sides
An arXiv preprint measured Model Context Protocol (MCP) tools that delete data or move money. It estimates 90% of destructive tools in a 1,000-repository sample of the official MCP registry have no server-side confirmation. Of 41 open-source hosts and agent SDKs, 32 have no approval step by default.
Why it matters
Agents can run destructive tool calls without a human confirming them, so an injected or mistaken instruction can delete data or move money. The study found this gap on both the server side and the host side, so a single control may not close it.
Do this
- Inventory MCP servers and tools in use, and flag any tool that deletes, modifies, or transfers data or funds.
- Require host-side approval for every tool not annotated read-only or non-destructive, and verify this setting in each agent host and SDK.
- Do not rely on a model-settable flag as the approval control for destructive actions; require a human decision outside the model.
- Where an MCP server uses elicitation, confirm it fails closed when the host cannot relay the question.
- Log and alert on destructive tool calls so unexpected deletions or transfers are detected.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“an estimated 90% of destructive tools (95% CI 82-95) have no server-side confirmation”
“Of 41 popular open-source hosts and agent SDKs, 32 have no approval step by default, including 16 of 25 applications.”
“Pairing sampled destructive tools flagged by a rule-based extractor with these applications, 60.7% of pairs have no enforced human approval by default.”
Sources
Affected / related
MCP elicitation, Destructive tool approval, Model Context Protocol
Published 2026-11-02 04:57 UTC · updated 2026-11-08 17:54 UTC · AI-assisted summary, checked against quoted sources · How we score