A regulated multi-agent assistant can read user input, call internal tools, receive tool responses, and produce user-facing answers. Security review found filtering only on the final answer, allowing jailbreak text in inputs and sensitive data in tool responses to influence intermediate reasoning and tool selection. Which guardrail strategy best addresses the gap?
Select an answer to reveal the explanation.
Short Explanation
Guardrails that only polish the final answer miss the dangerous middle: poisoned inputs, bad tool calls, and sensitive tool responses feeding the next thought. You need guardrails that inspect and enforce policy at input, tool call, tool response, and output stages. Output-only filtering, no guardrails, and user self-reporting fail regulated multi-agent security.
Full Explanation
A secure multi-agent solution requires a multi-intervention guardrail strategy covering user inputs, tool calls, tool responses, and outputs. Intermediate stages can leak data or steer tools even when the final answer is filtered.
The single output filter on the final message only is incorrect because output-only filtering leaves input injection and tool-path abuse unmitigated.
Disabling all guardrails in production to maximize model creativity is incorrect because it increases policy and safety risk in regulated settings.
Relying solely on users to self-report when they successfully jailbreak the system is incorrect because user self-reporting is neither preventive nor reliable control.
Exam focus: guardrails are multi-stage for agentic systems, not a single exit filter.