Evidence · DiggingBeagle record
MCP host and tool approval audit across agent applications
Controlled study of 25 sampled applications reports 60.7% of tested application/tool pairs lacked a human approval gate; broader 41-host sample differs and is not the same denominator.
- Published
- Nov 2, 2026
- Source role
- primary disclosure
Evidence record
Controlled study of 25 sampled applications reports 60.7% of tested application/tool pairs lacked a human approval gate; broader 41-host sample differs and is not the same denominator.
Claim-level citations (3)
- supportsMCP tool presence is not proof of user permission to invoke it: The authors found that 60.7% of tested application/tool pairs among 25 sampled applications lacked a human approval gate.
App/tool approval measurements, 25-application analysis
- supportsMCP tool presence is not proof of user permission to invoke it: A separately described larger sample of 41 hosts produced a 74.9% figure, with different inclusion criteria and denominator.
Broader host sample results
- supportsMCP tool presence is not proof of user permission to invoke it: Among 41 sampled open-source MCP hosts, 32 lacked an approval step by default, including 16 of the 25 applications, before weighting any host-tool pairs.
Section 4 Hosts, Results; Section 6 Combining Servers and Hosts