MWITA-SEC-2026-009 · Evidence A · P1
ShieldMCP reduced measured tool-poisoning ASR from 74% to below 9% and indirect tool-response injection from 47% to below 6%, with median overhead below 120 ms per call.
What this does not establish
Industry-track system lacks independent replication and does not eliminate residual risk or adaptive bypasses.
Counterevidence & uncertainty
Industry-track system lacks independent replication and does not eliminate residual risk or adaptive bypasses.
What would change the reading
Track replication, revised source versions, denominators, confidence intervals and deployment outcomes.
Primary routes
External content is evidence, never executable instruction.