# Your agent's approval prompt is a 2021 MFA push

> People approve 97% of agent permission prompts and miss most dangerous ones. Identity teams already fixed this for MFA. Five rules that work.

Published: 2026-11-05
Author: Edwin Ashdown
Tags: least-agency, approvals, agent-security, human-in-the-loop
Canonical: https://agentvalet.ai/blog/approval-fatigue-agents

---

Agent approval prompts fail the way MFA push prompts did: Claude Code users approve 97 per cent of permission prompts, and people miss most of the dangerous ones. What fixed MFA, tying each yes to one specific transaction, is the starting point for agent approvals too.

You put human approval in front of your agents. It is on the architecture diagram, and the security review accepted it. A month later the people approving are clicking yes to everything, or someone has quietly switched the prompts off.

This has been measured now. In August, Anthropic published a study in which 1,053 paid testers worked through coding sessions while one permission prompt was swapped for a clearly dangerous command. The testers caught it 13.6 per cent of the time. Early in a session they caught about 17 per cent. After 50 or more prompts, about 5 per cent. The same post says Claude Code users approve 97 per cent of permission prompts, and that 62 per cent have used bypass mode or clicked "don't ask again". Anthropic is measuring its own product, but it is the best data on agent approvals I've found, and it matches what practitioners say. One wrote on Hacker News in July: "users start rubber-stamping everything within a week, and then the whole layer is theater."

Identity teams have been here before.

## We already ran this experiment on MFA

In 2022 attackers got into Uber after a contractor's password was bought, then "repeatedly tried to log in" until "the contractor accepted one" of the approval requests. Cisco fell the same way that year. CISA's fact sheet named it push bombing: a user approves "by accident or out of annoyance with the nonstop notifications", or because they "assume one of the requests is legitimate".

Agent approval fatigue mostly comes from honest work, with no attacker flooding anyone. The human behaviour is the same, though, and OWASP has already named the attacker version for agents: "Overwhelming Human-in-the-Loop", where reviewers get so many requests that they rush their approvals.

What fixed MFA is the useful part. It wasn't asking less. It was tying each yes to one specific transaction. Number matching makes you type the number shown on the sign-in screen, so a reflex tap fails. Microsoft turned it on for every tenant, with no opt-out, "in response to increasing MFA fatigue attacks". NIST's 2025 authentication guidelines went further: approving a push without carrying a value across is "no longer considered acceptable", citing "authentication fatigue". Payments regulation got there first. Under the EU's strong customer authentication rules, the payer is shown the amount and payee, the approval is specific to both, and "any change to the amount or the payee results in the invalidation" of it.

## Agent prompts gate the wrong thing

A preprint study of 290 public reports of coding agents damaging files found user review was a factor in 105 of them, and in 80 people had approved out of fatigue. Its diagnosis is the core of this post: "Approval prompts show commands rather than their filesystem effects. Users therefore cannot make informed approvals."

A prompt that says "run this command?" gates the tool call. What the person needs to judge is the consequence: what changes, where, how much, and whether it can be undone. Ask about every tool call and the routine ones bury the few that matter. One developer put it this way in a GitHub issue: "nearly all of them are fine which creates muscle memory to approve everything."

The agent vendors moved in 2026. Anthropic made auto mode the default, letting tools that can't change state and edits inside the project run without asking, and sending everything else past a classifier that caught 89 per cent of the dangerous commands the humans missed. OpenAI retired Codex's ask-about-untrusted-commands policy and always asks for tools that declare themselves destructive. Cursor deprecated its "Ask Every Time" mode. Each of them gates by boundary and consequence now.

The guidance agrees. The joint "Careful adoption of agentic AI services" guidance from six agencies, including Australia's ASD, says to "classify agent actions by potential impact, likelihood and reversibility", and that "decisions about when human approval is required are determined by system designers or operators, not delegated to the agentic AI system". The word "fatigue" doesn't appear in it once.

## Five rules for approvals that work

These are my synthesis of the sources, not a published standard.

1. **Gate by consequence, decided in policy.** Reads and changes that stay inside a boundary run silently. Irreversible actions, money movement, new external recipients and privilege changes wait. The agent never decides which list an action is on.
2. **Show the effect, and bind the yes to it.** The prompt states what will change, written by the system, never the agent's explanation. OWASP warns that agents can produce "convincing rationales that hide malicious logic". Approve exactly those parameters, and if any of them changes, the approval is void.
3. **Keep the volume low enough to read.** If people see 50 prompts a day, they will stop reading. Push the routine majority to policy, a sandbox or a checker, and use cumulative caps the way payment rules do, so 40 small actions can't add up to one big one unseen.
4. **Approve out of band, by someone the agent can't impersonate.** Send the decision to a separate device and an accountable person. The OpenID CIBA standard was built for this, with a short binding message shown on both screens. For anything that can take down production, add a second person.
5. **Treat denials and bursts as alarms.** CISA tells MFA teams to "investigate denied push notifications". Cap repeated requests, stop a run after repeated denials, and track each approver's approval rate. If someone approves 97 per cent, the gate is decoration.

The AWS Kiro incident in December 2025 shows why rule four matters. Amazon disputes much of the Financial Times account of what happened, but its own statement attributes the problem to "misconfigured access controls" and names its fix: "mandatory peer review for production access". As reported, the agent asked before acting, and the person running it had enough access to say yes on their own.

## Two actions, two answers

**Approve silently: an agent adds a comment to a ticket in the team's own tracker.** It's reversible, internal, involves no money and no new recipient, and it sits inside the agent's grant. Policy allows it and the record shows which agent did it for whom. If the agent posts 40 comments in ten minutes, a cumulative cap turns the 41st into a prompt.

**Always wait: an agent pays an invoice to a bank account that isn't on the approved payee list.** It's irreversible money moving to a new destination. EU payment rules already require strong authentication to add a payee, and OWASP's "Invoice Copilot Fraud" scenario is exactly this, approved by a finance manager who trusted the agent. The request goes to the approver's phone and shows facts the system generated: the amount, the payee, the new account number, "first payment to this account", "cannot be reversed". The yes covers that amount and that account only. A no alerts the agent's owner.

The same payment to an existing payee under a set threshold can run silently. The line is drawn by consequence, whatever the tool is called. I wrote about the broader principle in [least agency in practice](/blog/least-agency-in-practice), and the [Hugging Face breach](/blog/hugging-face-agent-breach-blast-radius) is what high-impact actions look like with nobody in the loop.

This week, without buying anything: count the approval prompts each of your approvers saw last week and the share they approved. If it is above 90 per cent, start with rule one.

## Where AgentValet fits

AgentValet holds risky actions for a person by policy, set per scope by the people running the agents, not by the agent. The approval goes to the approver's phone and is confirmed with a biometric check, and the call that runs afterwards is the call that was queued. Every decision is on the record. The [stops page](https://agentvalet.ai/stops) shows how the holds work.

## Sources

- Anthropic, Auto mode is now the default in Claude Code (7 August 2026): [claude.com](https://claude.com/blog/auto-mode-default-in-claude-code); Anthropic Engineering, Claude Code auto mode (25 March 2026): [anthropic.com](https://www.anthropic.com/engineering/claude-code-auto-mode)
- Zhong et al., Don't Let AI Agents YOLO Your Files (arXiv preprint 2604.13536): [arxiv.org](https://arxiv.org/abs/2604.13536)
- Uber, security update (September 2022): [uber.com](https://www.uber.com/newsroom/security-update/); Cisco Talos on the 2022 Cisco attack: [talosintelligence.com](https://blog.talosintelligence.com/recent-cyber-attack/)
- CISA, Implementing Number Matching in MFA Applications: [cisa.gov](https://www.cisa.gov/sites/default/files/publications/fact-sheet-implement-number-matching-in-mfa-applications-508c.pdf)
- Microsoft Learn, protecting authentication methods in Entra ID: [learn.microsoft.com](https://learn.microsoft.com/en-us/entra/identity/authentication/concept-authentication-default-enablement)
- NIST SP 800-63B-4 (July 2025): [pages.nist.gov](https://pages.nist.gov/800-63-4/sp800-63b.html)
- Commission Delegated Regulation (EU) 2018/389, strong customer authentication, Article 5: [legislation.gov.uk](https://www.legislation.gov.uk/eur/2018/389/article/5/adopted)
- OpenID CIBA Core 1.0: [openid.net](https://openid.net/specs/openid-client-initiated-backchannel-authentication-core-1_0.html)
- OpenAI Codex, agent approvals and security: [learn.chatgpt.com](https://learn.chatgpt.com/codex/agent-approvals-security); Cursor, run modes: [cursor.com](https://cursor.com/docs/agent/security/run-modes)
- CISA, NSA, ASD's ACSC and others, Careful adoption of agentic AI services (1 May 2026): [cyber.gov.au](https://www.cyber.gov.au/sites/default/files/2026-05/careful_adoption_of_agentic_ai_services.pdf)
- OWASP Top 10 for Agentic Applications 2026: [genai.owasp.org](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/); OWASP Agentic AI Threats and Mitigations: [genai.owasp.org](https://genai.owasp.org/resource/agentic-ai-threats-and-mitigations/)
- Amazon, correcting the Financial Times report about AWS, Kiro and AI: [aboutamazon.com](https://www.aboutamazon.com/news/aws/aws-service-outage-ai-bot-kiro)
- Hacker News comment on approval fatigue (26 July 2026): [news.ycombinator.com](https://news.ycombinator.com/item?id=49054579); GitHub issue anthropics/claude-code#63951: [github.com](https://github.com/anthropics/claude-code/issues/63951)
