A regulated multi-agent assistant must block toxic user inputs, prevent disallowed tool calls, sanitize untrusted tool responses, and filter unsafe final answers. A single output filter has been bypassed by malicious tool-return content influencing the model. What guardrail strategy should you implement?
Select an answer to reveal the explanation.
Short Explanation
A is the multi-intervention model. Guardrails must cover user inputs, tool calls, tool responses, and outputs—because attacks move between stages. Trusting tool returns is how indirect injection wins. No guardrails and first-message-only policies leave later turns and tools unprotected. Design layered interventions in the orchestration middleware for regulated multi-agent systems.
Full Explanation
Correct Answer — A
Guardrail design requires a multi-intervention strategy across user inputs, tool calls, tool responses, and outputs. Single-stage output filters miss tool-mediated attacks; validating each stage closes those gaps.
Why B is wrong: Tool responses can carry indirect prompt injection and must not be fully trusted.
Why C is wrong: Regulated workloads need stronger, not weaker, controls.
Why D is wrong: Threats appear throughout the session, not only on the first message.
Exam tip: Four stages—input, tool call, tool response, output—all guarded.