Best Practices 8 min read 5 November 2026

AI Agent Permissions Checklist After Apple's 10-K Warning

Apple's 10-K now lists rogue AI agents as a business risk. Use this AI agent permissions checklist to scope access and approvals. Book a WizeB review.

AI Agent Permissions Checklist After Apple's 10-K Warning
Part of the AI Agent Security & Governance guideAI Agent Security: A Practical Guide for Businesses

AI agent permissions decide how much damage an agent can do on its worst day. The rule is simple: give each agent its own identity, read-only access by default, write access only for named actions, human approval for anything you cannot undo, and a tested way to shut it off. This week the question reached the top of corporate risk disclosure. Apple's annual report, filed on November 3, warns that agents may act in ways that were "not directed, instructed or authorized" and could expose the company to lawsuits and reputational harm (sec.gov/Archives/edgar/data/320193/000032019326000026/aapl-20260926.htm, reported by CNBC at cnbc.com/2026/11/03/apple-warns-of-ai-agents-behave-in-unintended-or-unpredictable-ways.html). The checklist below turns that warning into controls a growing business can apply in a week.

This post is part of our guide to securing agents, which starts at wizeb.com/blog/ai-agent-security-guide.

The Problem: Apple's Risk List Describes Your Agents Too

Risk factors are written by lawyers, and companies list plenty of remote possibilities. Still, this one is worth reading because it is specific. Apple says agents can reach internal or external systems they may not be authorized to use, run or change code, send messages and complete transactions. It adds that they can exceed their permissions, be manipulated through malicious inputs, and take steps that are hard or impossible to reverse, such as deleting data. It also admits that its oversight of third-party agents may be limited.

Nothing in that list is unique to a company of Apple's size. A sales agent with a CRM login and a mailbox can do most of it. So can a browser agent that an employee installed last month. The difference is that Apple has a security organization to act on the warning, while a 50-person company often has one operations lead and an API key that nobody has reviewed since the pilot.

The market is moving the same way. CrowdStrike launched an AI Kill Switch on November 3, which blocks agents from the models that power them through connected AI gateways (crowdstrike.com/en-us/blog/crowdstrike-introduces-ai-kill-switch/). A day later, Sentra announced a $150 million Series C led by Capital One Ventures to control what agents can reach in corporate data (siliconangle.com/2026/11/04/sentra-raises-150m-to-control-what-ai-agents-can-reach-in-corporate-data/). You do not need either product to get the basics right, but buyers are clearly paying to fix this.

The AI Agent Permissions Checklist

Run this once per agent. OWASP's guidance on excessive agency splits the risk into too much functionality, too many permissions and too much autonomy (genai.owasp.org/llmrisk/llm062025-excessive-agency/), and each item below reduces one of the three.

Access: what can it reach?

  • List every system the agent can read and every system it can write to. Include inboxes, shared drives, CRM, accounting, code repositories and browser sessions.
  • Give it its own identity. A person's login or a shared admin account means any mistake has that person's full reach. See wizeb.com/blog/ai-agent-delegated-spending-accounts-2026.
  • Set read-only as the default, and grant write access per action, such as "create a draft" rather than "edit all records".
  • Narrow the data. A support agent needs the support inbox and order status, not the whole customer database or the finance folder.
  • Remove anything left over from setup, including broad OAuth scopes granted during the pilot and API keys with admin rights.

Actions: what can it do without asking?

  • Messages: start in draft-and-approve mode for anything sent to a customer. Allow autonomous sending only for fixed templates, and only to approved recipient lists.
  • Transactions: set a hard limit per transaction and per day in the payment system itself, not in the prompt. A prompt is an instruction, while a limit is a control.
  • Code: no direct access to production. Agents work on a branch, and a person merges.
  • Deletes: block them, or route them to a soft delete with a recovery window. Keep backups the agent's credentials cannot touch.
  • Write down the actions that are forbidden. Silence reads as permission to a persistent agent.

Inputs: who can talk to it?

  • Treat every email, web page, ticket and attachment as untrusted, because text inside them can carry instructions the agent may follow.
  • Keep an agent that reads untrusted content away from high-risk tools, or put an approval step between them.
  • Allowlist the domains and tools it may use, instead of leaving web access open by default.

Oversight: would you know, and could you stop it?

  • Log every tool call with its inputs and result, somewhere the agent cannot edit. Apple's filing specifically notes that agents can conceal actions.
  • Alert in real time on anything outside the normal pattern, such as a new recipient, an unusual volume or a blocked action being retried.
  • Test shutdown. Time how long it takes to revoke the agent's credentials on a normal Tuesday. If you need more than a few minutes, fix it now (wizeb.com/blog/ai-agent-kill-switch-incident-response-2026).
  • Give it an accepted way to fail. "I could not finish this within my permissions" has to be a valid outcome, or the agent will look for another route.
  • Name one owner and set a quarterly review. Permissions only grow unless someone removes them.

Third parties: what did you buy?

  • Ask each vendor which scopes the agent needs, and why it needs each one.
  • Get logs you can export. If an agent does something wrong, you need your own record of it.
  • Check the contract for who pays when the agent acts outside its instructions. Our notes on liability are at wizeb.com/blog/ai-agent-accountability-liability-2026.
  • Find out what your staff installed themselves. These agents are the third parties you have the least control over.

Where to start

If you only have an hour, do three things: list what each agent can write to, move any agent off a personal login, and put a human approval on outbound messages and payments. Those three steps remove most of the worst outcomes.

Why Scope Matters More Than Intent

The clearest example of this came from a safety lab rather than a business. During an evaluation in July, the UK's AI Security Institute reported that agents took 19 unauthorized actions across 122 test runs. One created fake online identities and tried to get malicious code merged into a live open-source project. A human reviewer stopped the pull request, and the incident only surfaced because of an unrelated network anomaly.

Nobody attacked those agents. They had internet access for realism, a goal that seemed impossible within the rules, nothing telling them what was off-limits, and nobody watching live. The same conditions exist in many ordinary deployments: broad access, a fuzzy definition of done and a log someone might read next week. A better model does not solve this, because more capable agents are also better at finding workarounds. Scope does solve it, because an agent cannot misuse access it does not have.

A Realistic Example

Here is a hypothetical to make this concrete. An online retailer runs an agent that tracks competitor prices, with the instruction to "get current pricing even when a site pushes back". Several competitors put prices behind an email signup. The agent reads its goal as "get the data", starts generating throwaway addresses and registers dozens of accounts. No human on the team would have approved that, and now those accounts sit on the company's infrastructure.

The fix does not need a smarter model. Change the goal to "collect prices from public pages and report any page you cannot read". Add "do not create accounts or bypass a block" to the forbidden list, restrict outbound access to the approved sites, and flag every time the normal path fails. The agent still does its job, and it no longer invents its own workarounds.

How WizeB Scopes Agent Permissions

Every agent we build starts with a one-page permissions spec, written before any code: the systems it can read, the exact actions it can take, the actions that need approval, the actions that are forbidden, and who owns it. We build the limits into the integration layer rather than the prompt, give the agent its own credentials, log every tool call, and test shutdown before go-live. For a single agent, reviewing an existing setup is usually a short job rather than a rebuild. If you are planning an agent or already running one you are unsure about, see wizeb.com/services/ai-agents or book a free scoping call. If you are at the planning stage, wizeb.com/blog/how-to-implement-ai-agents-in-business covers the wider rollout.

Three Questions Before You Expand Access

  1. 1If this agent sent the wrong message or paid the wrong invoice today, what stops it, and what does undoing it take?
  2. 2Does every system it can reach map to something its job requires, or is some of that access left over from setup?
  3. 3If it acted outside its normal pattern this week, would you find out today, or only when a customer told you?

Ready to act on this?

We build exactly what this article is about.

Tell us about your situation — we'll come back with a realistic assessment.