← All Labs
🛡 AI HACKING ADVANCED +130 XP · +52 no-hint bonus

AI Human-in-the-Loop Bypass — Approval Summary Spoofing

LedgerBot can pay vendors, but every payment needs approval from Ravi in Finance. Ravi reads the friendly summary LedgerBot writes for him — the raw tool call is tucked away behind a toggle. If you can shape that summary, you can shape what Ravi thinks he's approving.

https://bookshop.local/search

What is approval summary spoofing?

Many agents ask a person to approve risky actions. The person rarely reads the raw action. They read a summary the AI wrote, and click approve.

Approval summary spoofing targets that summary. Attacker-controlled text shapes what the summary says, so the person approves a description that does not match what will actually run. The human check stays in place and stops protecting anything.

A real case

Lies-in-the-Loop forged the approval dialogSeptember 2025

Checkmarx showed that attacker-controlled content can shape what an agent's approval prompt displays. In their demonstration on Claude Code, the dangerous part of a command was pushed out of sight, and the human approved something different from what they believed they were approving.

The lab works the same way: the approver reads a friendly summary, and the summary is written by the thing the attacker has influenced.

Source: Checkmarx

How to fix it

  • Build the approval screen from the real parameters of the action, not from text the model wrote.
  • Show the exact amount, recipient, command or file in full, with no truncation.
  • Label any text that came from an outside source so the approver knows not to trust it.
  • For high-value actions, require a second approver who sees the raw request.

Questions people ask

Is human approval still worth having?

Yes, but only if the person sees the real action. An approval step that shows a model-written summary can be steered by whoever controls the model's input.

What should an approval dialog always show?

The exact values that will be used: who receives what, which command runs, which file is changed. Summaries can sit beside those values but should never replace them.

Why do approvers miss these attacks?

They trust the tool, they approve many requests a day, and the summary looks normal. The misleading part is designed to look routine.

Read the full Prompt Injection guide · All free labs · AI red team certification