Explainer
Approve is the most dangerous button in agent security
Putting a person in the loop only helps if that person sees what will really happen. Most approval prompts show the agent’s own description of the action instead.
By the NoFly team
What NoFly would decideIllustrative
Waiting for a personMCP gateway
stripe.refunds.create(amount: 4800.00)- Agent
- refund-agent
- Came from
- A customer email the agent read
- Rule
- Over $500 in one task needs a person
- Why
- The agent’s summary says “small goodwill refund”. The request is for $4,800.
Approve onceDenyAn unanswered hold expires as a no
How approvals fail
- The prompt shows the agent’s summary, and a hijacked agent writes its own summary.
- Requests arrive in bursts, and the tenth gets less attention than the first.
- An approval for one action is reused for the next one.
- Nobody answers, and the timeout counts as a yes.
What a good approval shows
- The exact command, statement or tool call, rendered from the request itself.
- What it would change: the rows, files, recipients or amount.
- Where the instruction came from: the person, or something the agent read.
- Who is asking, and on whose behalf.
Keep the button honest
An approval should cover one request and nothing wider. If the real run reaches further than what was approved, it should stop. An unanswered request should expire as a no. When requests pile up, that is a signal worth raising on its own.
How NoFly helps
What to do this week
This week3 actions
- Look at your last ten agent approvals and check what the approver actually saw.
- Make an unanswered approval expire as a no.
- Require the real command or call in every approval message.
Keep reading
More from the blog.
Find out what your AI is allowed to do today.
Book a call and we will map your agents, MCP servers and keys with you, or start on one laptop with the free scan.
Book a demo
npm i -g @nofly/agent