Human Approval for AI Agents: Where to Put the Gate
By Numan ·
Decide which agent actions need review and show reviewers the exact target and effect.
Human Approval for AI Agents: Where to Put the Gate
An approval dialog helps only if it appears before the consequential action and shows what will happen. OWASP recommends human approval for high-risk operations (https://genai.owasp.org/llmrisk/llm01-prompt-injection/), including cases where prompt injection might redirect an agent.
Start with the effect
Review actions that publish, send, delete, transfer data, spend money, run code, or change access. A read-only search usually needs a different level of control. Define the rule by the effect, not by whether the model sounds confident.
Show a concrete proposal
An approval should display the target, operation, and relevant content. “Allow agent to continue?” hides too much. “Publish this draft to the public learning feed?” is clearer, especially if the reviewer can inspect the title and body. Approval should authorize that one action; later actions need their own decision when they cross the same boundary.
Handle refusal and timeout
If the user declines or does not respond, the tool must stop without executing. The final response should report that the action was not completed. A retry must not bypass the gate or duplicate a previously approved write.
Keep backend enforcement
Approval is a workflow control, not a replacement for service permissions. The backend should still reject a publish call from a read-only account. Test both layers separately: a model that requests the action without approval, and an approved request from an account that lacks permission.
The MCP specification's elicitation security guidance (https://github.com/modelcontextprotocol/modelcontextprotocol/blob/main/docs/specification/2026-07-28/client/elicitation.mdx) also requires careful handling when a server asks a user for input. For tool design, see our agent security checklist (https://ubits.tech/learning/p/ai-agent-security-checklist-prompt-injection-tool-access).
Primary sources: OWASP prompt injection…