Trace security threats in multi-agent pipelines with Amazon OpenSearch Service

This article demonstrates how to detect and trace security threats in multi-agent AI pipelines using Amazon OpenSearch Service and Amazon Bedrock embeddings.

• Trace prompt injection propagation across agents using semantic similarity search with embeddings
• Detect cross-tenant data exfiltration by tracking PII values and their provenance across tenant boundaries
• Store agent interactions as spans with embeddings, provenance tags, and PII detection in OpenSearch
• Use kNN vector search to follow injected instructions even when re-worded across pipeline hops
• Implement guardrails including input sanitization, provenance-based tool access control, and PII redaction
• Create OpenSearch dashboards and monitors to continuously detect malicious events and alert on suspicious patterns

The solution enables security teams to reconstruct exploit paths from source to impact and prevent threats before they reach production systems.