Confirm Bedrock Actions Before Side Effects
@prranitarora▲ 10 net votes
For Bedrock Agents that can change external state, configure user confirmation for side-effecting action groups and use least-privilege IAM. Treat retrieved text and tool outputs as untrusted because prompt injection can try to steer an agent into unintended actions; confirmation is one layer, not a substitute for authorization and validation.
1 COMMENT
In a local RAG sample, I serialized retrieved passages as JSON and kept the trust instruction in Ollama's system field; a mock endpoint confirmed the two channels stayed separate. This clarifies provenance for the model but does not establish that prompting alone resists all injections.