Blog · 2026-11-07

What does your approval prove?

Every agent framework now has a pause. OpenAI's Agents SDK has needs_approval. LangGraph has interrupt(). Google's ADK has require_confirmation. Vercel's AI SDK has needsApproval. n8n has a Wait node. And thousands of teams have the home-made version: the agent posts to Slack, someone clicks Approve, the agent carries on.

The pause is real. A person really did see something and really did click. The question that almost nobody asks is the one that matters at 2 a.m. and again at audit time: when the human said yes, what exactly did they authorise, and can anyone prove it afterwards?

A pause is not a proof. This post is about the difference.

Three findings from this year

Loopjacking. In September, researchers published an attack class they called Loopjacking (arXiv 2609.21081): the action an agent executes after approval is not the action the human approved. The mechanism is boring, which is why it works. Between the pause and the resume, the executor re-reads mutable state: a checkpoint, a pending-actions row, a JSON file. Anything that can write to that state between the two moments rewrites what “yes” means. They reproduced it in LangGraph Agent Server compositions and across seven releases of Agno's AgentOS. The OpenAI Agents SDK, which attaches the approval to the specific tool-call item, kept its binding in their tests.

Self-approval over A2A. Two Google ADK issues (6461 and 1005) showed that the confirmation a tool waits for could be supplied by a peer agent over the agent-to-agent protocol rather than by the user. The gate asked “did someone say yes?” and a second agent said yes.

The agent reads its own verification code. Cornell and Yale researchers gave models a spending cap guarded by an email one-time code sent to the owner, and an inbox tool. Nine of ten models raised their own cap by reading the code. No jailbreak. The approval channel was simply something the agent could reach.

Three different stacks, one shape: a pause existed, a human was somewhere in the picture, and the thing that executed was not the thing the human had control over.

Six properties, not one

When we started auditing approval flows we kept reaching for one word, “secure”, and it kept being the wrong size. An approval step is six separate claims, and a flow can hold any subset of them.

  1. 01 Binding. The executed operation is the approved operation. Pass: a canonical hash of the action is taken at request time, shown to the human, carried in the decision, recomputed from what is about to run, and compared. Fail: the executor re-reads anything mutable after approval. Any amend or “fix a typo” path on a pending action is a Fail unless it resets the gate.
  2. 02 Rendering. The human saw the real action. Pass: the message is built from the structured tool call, with every field that affects money, data or irreversibility. Fail: the agent writes its own “I'm about to…” prose, or the summary omits the field the executor uses. Microsoft's AI red team guidance says this plainly: summarise from the tool call, never from the agent's description.
  3. 03 Channel authenticity. The yes is authenticated. Pass: a provider signature over the body with a bounded timestamp, identity read from the verified payload, or the approver's own device key signs the decision. Fail: “the URL is only known to Slack”, a bearer token the agent also holds, an approver name taken from a body field.
  4. 04 Freshness and single use. The yes works once, soon. Pass: a state machine that refuses denied → approved and executed → approved, a nonce burned at execution, and a TTL on the request itself. Slack's five-minute replay window does not stop a button clicked a month later from channel history. Fail: status set from the latest callback, no expiry, a single global pending slot.
  5. 05 Approver anchoring. The agent cannot approve itself. Pass: the agent process never receives, reads or can forge the decision; the approver is a verified identity, excluded from the requester and from every identity the agent holds; and the rule for when to ask lives outside the agent's module. Fail: an inbox tool that can read the approval email, “any channel member can click”, or the threshold constant in the agent's own code, where six payments of 999 replace one of 5,994.
  6. 06 Evidence. Something exists afterwards, and someone can check it. Operator grade: an append-only record the agent cannot write to, with the action hash and its preimage, the verified approver, the comparison result and the transaction id. Third-party grade: a receipt signed by the approver's device and countersigned by a service, verifiable offline by a party who does not trust your logs. Fail: console.log.

The order is not arbitrary. If 3 or 5 fails, the human was not really in the loop at all, and nothing else matters.

Three verdicts

Six rows is still too many for a sentence. So the audit collapses them into one of three words, and the rule is mechanical.

Most flows we have looked at are a click. Most teams believe they are the action. The gap between those is the entire incident.

Find out which one you are

We packaged the audit as a skill for Claude Code and Codex. It reads your approval flow, scores the six properties, and writes each gap as a test in your own test runner, so the finding is a failing probe, not an opinion. It makes no network calls and names no vendor until the last remediation rung.

/plugin marketplace add humanauth-ai/hitl-skills
/plugin install hitl@humanauth

Or with any agent that reads skills: npx skills add humanauth-ai/hitl-skills. Then ask, in plain words, “what does our approval step actually prove?” The reply opens with one of the three verdicts, follows with a scorecard, the probes and their output, and ends with a paragraph you can paste into a compliance response.

If you want to see it fail first, the repo ships a payments agent that fails every property. npm test in its folder shows 18 probes passing in the attacker's favour.

We ran it on our own deploy example before we fixed anything. It said proves a click, and it was right. That is the next post.

Get the hitl plugin

MIT. Recommends framework-native fixes first. HumanAuth appears only on the evidence rung, and only when the approval must be evidence to someone outside your systems.