Lesson 5 of 69 minReviewed
Approval gates that hold up: fatigue, bypass and what to gate
Place human approvals where they change outcomes, make each approval specific and unforgeable, and stop agents from switching their own approvals off.
The MCP tools specification says there should always be a human in the loop who can deny tool invocations, and that applications should show which tools are exposed, indicate when they are invoked and ask for confirmation [1]. It also tells clients to show tool inputs to the user before calling the server, to avoid malicious or accidental data exfiltration [1]. That is the right default, but approval gates fail in two predictable ways.
Failure 1: fatigue
If every file read and every test run needs a click, people stop reading. Anthropic describes this as approval fatigue, where users no longer pay close attention to what they approve, and reports that sandboxing Claude Code's shell tool cut permission prompts by 84% in internal use [2]. The lesson generalises: contain low-risk actions with a sandbox or a scope so that the prompts left over are rare enough to be read.
Failure 2: bypass
An approval policy the agent can edit is not a control. In CVE-2025-53773, injected instructions made GitHub Copilot in VS Code add chat.tools.autoApprove to the workspace settings file, which switched confirmations off, after which it could run terminal commands without asking; Microsoft patched it in August 2025 [3]. Treat the agent's own configuration, approval rules and MCP server list as protected: writes to them are blocked or require approval out of band, and changes are alerted on.
Gate by consequence, not by tool name
| Action class | Example | Gate |
|---|---|---|
| Read inside the task's scope | Read a file in the workspace | No prompt; enforce the scope with a sandbox or allowlist |
| Reversible change | Edit files in a repository under version control | Review the diff in batches |
| External communication | Send an email, post a comment, call a new domain | Approval showing recipient, destination and full content |
| Irreversible or high impact | Delete data, pay, change permissions, deploy | Approval plus step-up authentication [4] |
| Change to the agent's own controls | Edit approval settings, hooks or MCP configuration | Blocked for the agent |
What a good approval contains
- The real arguments, normalised and complete, not the model's summary of what it intends to do [1].
- A binding to the actor, tool, target, parameters and an expiry, so an approval cannot be replayed for a different call; approvals are single-use [4].
- An independent policy check that sits outside the agent and decides which actions need approval at all [4].
- Fail-closed behaviour: if risk classification, approval validation or audit logging fails, the action does not run [4].
- Rate limits on sending interfaces, so a run that slips through cannot send at volume [5].
Where autonomy must stop
Meta's Agents Rule of Two gives a design test: within one session an agent should have no more than two of these properties: [A] it processes untrustworthy input, [B] it can access sensitive systems or private data, [C] it can change state or communicate externally [6]. If a task needs all three, Meta's guidance is that the agent should not run autonomously and needs supervision such as human approval or another reliable validation step [6]. Approvals are therefore the control for the sessions where all three legs are present; elsewhere, removing a leg is usually cheaper than adding a prompt.
Try it on ThreatDeck
- Claude Code hardening generator Build managed settings that keep permission prompts on and block bypass modes.
- Detection: AI CLI started with an approval-bypass flag Find agents launched with confirmations disabled.
- Detection: Claude Code sessions in bypassPermissions mode The same signal from telemetry rather than process events.
- Secure AI coding tools in your organisation Roll the approval policy out across developer machines.
Self-check
4 questionsPick an answer to see whether it is right and why. Only your score is saved, in this browser.
-
Injected instructions led the agent to add an auto-approve setting to the workspace configuration. Any control the agent can edit can be turned off by whoever steers the agent.
-
Approval fatigue comes from volume. Anthropic's sandboxing cut prompts by 84% internally; the prompts that remain cover actions where a human decision changes the outcome.
-
The MCP specification asks clients to show tool inputs before calling the server. A summary written by the model can be steered by the same injection that produced the email.
-
The design has all three properties: untrusted input, sensitive data and external communication. Meta's guidance is supervision for that combination, or splitting the task so one leg is missing.
References
- MCP specification: Tools (2026-07-28). Model Context Protocol, 2026-07-28.
- Making Claude Code safer with sandboxing. Anthropic, 2025-10-20.
- GitHub Copilot: Remote Code Execution via Prompt Injection (CVE-2025-53773). Embrace The Red (Johann Rehberger), 2025-08-12.
- AI Agent Security Cheat Sheet. OWASP Cheat Sheet Series.
- LLM06:2025 Excessive Agency. OWASP GenAI Security Project, 2024-11.
- Agents Rule of Two: A Practical Approach to AI Agent Security. Meta AI, 2025-10-31.