Mmastodon TechnologyCybersecurity first seen 6 d ago, last 6 d ago, peak #1

Study finds most destructive MCP tool pairings lack human approval

Original: 60.7% of sampled pairings of a destructive MCP tool and an open-source app have no enforced human approval by default. S

A security analysis reports that 60.7% of sampled pairings between a destructive MCP tool and open-source applications have no enforced human approval by default. While both the server and host can request confirmation, in most cases neither does. Requiring hosts to relay elicitation raises the protected share only slightly, to 56.7%, leaving many risky actions able to run without a person signing off.

Why now: Researchers and developers are alarmed that agentic tool integrations can perform destructive actions with no default human-in-the-loop safeguard.

MCP (Model Context Protocol)open-source applicationsAnthropic

Open on mastodon →

Evidence

• 60.7% of sampled pairings of a destructive MCP tool and an open-source app have no enforced human approval by default. Server and host can each ask. Often neither does. Every host relaying elicitation only gets that to 56.7%. Hosts prompting on any tool not annotated read-only… · shellonline · 2

API: https://socialmediatrends-api.osmike.com/v1/trends/4880124