Trace security threats in multi-agent pipelines with Amazon OpenSearch Service
In this post, you learn how to trace two invisible threats in a multi-agent pipeline, prompt injection propagation and cross-tenant data exfiltration, using Amazon OpenSearch Service and Amazon Bedrock.
Multi-agent systems have interactions that traditional security tools can’t see. A prompt injection buried in an external document can hijack a downstream agent that has tool access. Sensitive data can cross a tenant boundary without tripping a single alert. The damage happens in the traffic between agents, and standard logging can’t reconstruct the path. Prompt injection is LLM01 in the Open…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.