[Feature]: Redact agent and MCP secrets from Agentcore output

[jeromevdl]

Description

AgentCore injects agent credentials and MCP secrets into CLI environment variables. Agent or tool output may accidentally expose these values through stdout or stderr.

Add per-invocation secret redaction before output reaches:

• CloudWatch/container logs
• DynamoDB execution output
• WebSocket clients
• Error messages and captured stderr

Requirements:

• Redact resolved agent credentials and MCP secret values.
• Handle secrets split across stream chunks.
• Recursively redact structured output.
• Ensure stdio: inherit cannot bypass redaction.
• Add CloudWatch Logs data-protection policies as defense in depth.
• Add tests for stdout, stderr, structured output, and chunk boundaries.

Redaction protects against accidental disclosure. It is not a security boundary against deliberate encoding or exfiltration by an unrestricted agent.

Use case

Secure logs

Area

Backend (Lambda)

Additional context

No response