MWITA-SEC-2026-003 · Evidence A · P1
PoisonedRAG reached up to 90% attack success by inserting five malicious texts per target query into knowledge bases containing millions of texts.
What this does not establish
Assumes corpus write access; answer corruption is not model-weight poisoning or successful tool execution.
Counterevidence & uncertainty
Assumes corpus write access; answer corruption is not model-weight poisoning or successful tool execution.
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.