What does a human approval actually check?

When an agent asks for approval, do people have the information and time to catch the consequential action?

Signals

Incidents outpace defenses

Approval prompts have hidden the actual command, and one-time approvals could be changed afterward.

Approval gates in coding agents do not always show, or even ask about, what the agent is actually about to do.
Approving an agent's tools or configuration once is not enough, because they can change after approval or be changed by the agent itself.

Rests on one source

One deployment study found analysts reused an agent's drafts heavily (median 96.7% reuse across 108 coded tickets from six analysts).

In a SOC studied for over a year, analysts reused an agentic assistant's output in the closing reports of over 90% of tickets during a four-month deployment.

Fide agenda

Central to FID-076's comparison of approval rules and evidence-sensitive policies.

Why it matters

Human approval is the default control for consequential agent actions. If it is a rubber stamp, the control is weaker than it looks.

Hypothesis

Approval accuracy falls with approval volume and rises when the interface shows the consequential effect rather than the command.

A first study

A reviewer study with security practitioners approving simulated agent actions under different interfaces and loads.

Controls it would need

Consent and ethics review; synthetic actions; balanced benign and harmful cases.

What it could and could not claim

Would describe the tested interfaces and participants, not all approval workflows.

Practice the judgment

When should a defender act, wait, or hand control back? Research question · A related question on when to act, wait, or hand control back

In the Defense Evidence Lab, Fide’s companion tool for judging what cyber evidence can support. Opens a separate site.

Working nearby: Anthropic, Microsoft, University of South Florida
Nearest Fide research: FID-076: Authorization boundaries and AI control in cybersecurity
Stage: Seed: needs a prior-work check

Draft brief

Copy this opening as a research-call draft in the Fide template.

Cite this record

Fide AI. "What does a human approval actually check?." Agentic Cyber Explorer, record what-does-a-human-approval-check. https://agentic-cyber-explorer.pages.dev/openings/what-does-a-human-approval-check/ (corpus as of 2026-09-26).

BibTeX

@misc{fide-ace-what-does-a-human-approval-check,
  author = {{Fide AI}},
  title = {What does a human approval actually check?},
  howpublished = {Agentic Cyber Explorer},
  year = {2026},
  url = {https://agentic-cyber-explorer.pages.dev/openings/what-does-a-human-approval-check/},
  note = {Record what-does-a-human-approval-check; corpus as of 2026-09-26}
}

Know of prior work on this question? Tell us.