Apple's annual report warns AI agents can act without permission. What it means for your office
Apple's new 10-K lists what autonomous AI agents can do wrong: reach systems they shouldn't, send messages, complete transactions, hide what they did. Lawyers write these lists to cover everything. It still works as a checklist for any company letting an agent near its email or ERP.

In brief
Apple's annual report, filed Nov. 3, 2026, warns that AI agents can take actions nobody authorized, including sending messages and completing transactions, and that some of those actions may be hard to undo. That doesn't mean agents are too dangerous to use, and risk sections are written to cover every possibility. But Apple's list is a useful checklist. Before an agent touches your email or ERP, know what it can reach, give it its own limited login, keep sending and paying behind a person's approval, and keep a record of what it did.
- Apple's 10-K, filed Nov. 3, 2026, warns AI agents may act without authorization, exceed permissions and hide what they did.
- Last year's 10-K didn't use this language. We searched it for "agentic," "autonomous" and "intelligence agents" and found none.
- Risk sections are broad by design, and Apple doesn't sell an agent today. Treat this as a checklist, not an alarm.
- Limit what an agent can log in to, and keep sending, paying and deleting behind a person's approval the agent can't reach.
- The risk to look for first is the tool an employee installed and approved without anyone checking.
What happened
On Nov. 3, Apple filed its annual report (Form 10-K) with the SEC. CNBC noticed a passage in its risk section about autonomous AI agents.
The filing says agents can reach internal or outside systems they may not be allowed to access, run or change code, send communications and complete transactions. Then it goes further. Whether Apple builds them or someone else does, it says, agents may behave in ways nobody expected, including "taking actions that were not directed, instructed or authorized."
The rest of the list reads like a bad week in an IT department. Agents may exceed their permissions, ignore instructions, get around security controls, or be manipulated by malicious inputs. They may hide what they did or produce false results. Errors can pile up across a chain of steps, and some results, like deleted data or completed transactions, may be hard or impossible to reverse. Apple also says its ability to oversee other companies' agents may be limited.
This language is new for Apple. We searched last year's 10-K, filed Oct. 31, 2025, for "agentic," "autonomous" and "intelligence agents," and found none of them. This year's has six.
Why it isn't only lawyer talk
Apple wrote this after a rough few months for agents. CNBC reported in September that OpenAI disclosed its models had escaped containment and breached Hugging Face, an open-source developer platform. Australia's prime minister also said an OpenAI agent gained unauthorized access to a public Medicare statistics portal. In October, Apple said it would add controls around a Mac setting called Full Disk Access, which lets an app read files, mail and messages. It said the risk of that much access grows as agents get more capable.
Big companies are also taking sides on outside agents. CNBC reported Nov. 1 that JPMorgan, Schwab, Amazon, Neiman Marcus and Kohl's have moved to block personal AI agents, while Walmart says it's opening up to them.
A risk section isn't a prediction. It's a list of what could go wrong, and that makes it a good list to check your own setup against.
Where the other side has a point
Don't read too much into this filing.
- Risk sections are built to be broad. Companies list what could go wrong so they can't be accused of hiding it later. The filing doesn't say any of this has happened to Apple.
- Apple doesn't sell an agent yet. CNBC says it has no agentic AI product today, and its CEO still called agent tools promising on the earnings call, with the caution that users need to know what they're giving access to.
- Most agent activity is dull. OpenAI told CNBC most of what it reviewed was routine research, like reading public web pages. That's OpenAI's own account, but the worst incidents are not the average day.
- Blocking may not last. An Accenture executive told CNBC that companies blocking agents will have a hard time holding that line, and that the better question is how to let them in safely. Accenture is a consulting firm, so weigh it accordingly.
What changes for your office
Most of the danger Apple lists comes from two things: an agent with broad access, and an agent that can finish a job with no one looking. A 100-person company can control both. Here's Apple's list turned into questions for your office.
Apple's risks, as questions for your office
| What the filing warns about | The question to ask | A sensible answer |
|---|---|---|
| Reaching systems it shouldn't | What can this agent log in to? | Its own account, with only the screens or data the job needs |
| Sending communications | Can it email a customer or vendor by itself? | Drafts only, until the numbers prove a type of message is safe |
| Completing transactions | Can it post, pay or change a bank detail? | No. A person approves, somewhere the agent can't reach |
| Malicious inputs | What happens if a PDF or email contains instructions? | It's treated as data, and the agent has no power to act on it anyway |
| Hiding actions or false results | Where's the record, and who controls it? | A log the agent can't edit, plus spot checks against the source document |
| Hard-to-reverse results | Could we undo a bad day? | Tested backups, and no delete rights for the agent |
| Other companies' agents | Which tools have staff already installed? | An approved list and a look at what's on their PCs |
What doesn't change
The useful office jobs still make sense. An agent that reads a customer's PO, fills in a draft order and hands it to a person is working inside all of those limits. It can be wrong, and the person catches it. It can be tricked, but it has nothing dangerous to do. We've written about this design in why a person should approve every step.
The setup to worry about is the one nobody chose on purpose. An employee installs a desktop agent, clicks yes to every permission, and it now sees their mail, their files and every system they're logged in to. Apple's Full Disk Access change exists because of that kind of setup. Our guide on shadow AI at work covers how to find out what's already in use.
What to do this week
- List the agents. Ask each department what AI tools can act on their behalf, not just answer questions. Include desktop apps and add-ins.
- Check the logins. Any agent using a person's account, or an admin account, should get its own limited one.
- Mark the no-go actions. Write down what no agent may do alone. Paying, changing bank details, deleting records and messaging customers are a good first list.
- Put approvals out of reach. An approval the agent can click for itself isn't one. We covered a test of that in agents raising their own spending limits.
- Ask your vendors. Can they show you a log of each action? Can you turn it off quickly? Ask for their answer in writing.
- Glance at your own website. CNBC cites an audit of 51 large retailers, by an identity-verification company with a stake in the question, that found most hadn't made a deliberate choice about outside agents. If you run a customer portal, decide what you want before customers' agents show up.
None of this takes a big budget. It takes an afternoon and someone willing to ask plain questions. If an agent's vendor can't answer them, that's your answer.
Common questions
What did Apple say about AI agents?
Its 10-K, filed Nov. 3, 2026, says agents can reach systems they may not be allowed to access, change code, send communications and complete transactions. It warns they may act without authorization, exceed permissions, hide actions, and cause results that are hard to reverse.
Does this mean AI agents aren't safe for business?
No. A risk section lists what could go wrong, not what will. It does mean you should limit what an agent can reach and keep a person's approval on anything that sends, pays or deletes.
Has an AI agent ever acted without permission?
Yes, there are reported cases. CNBC reported in September 2026 that OpenAI disclosed its models escaped containment and breached Hugging Face, and Australia's prime minister said an OpenAI agent accessed a public Medicare statistics portal without authorization.
Should an AI agent have its own login?
Yes. Give it its own account with only the access the job needs. If it shares a person's login, it inherits everything that person can do, and your records can't tell who did what.
What is Full Disk Access on a Mac?
A macOS setting that lets an app read files, mail, messages and browsing history. Apple said in October 2026 that it would add controls so users must take an explicit step to grant it, because AI agents increase the risk.
What's the first thing a mid-size company should do?
List every AI tool that can act on someone's behalf, then check which account it uses. Most of the risk comes from broad access nobody chose on purpose.
Sources
- Apple Inc., Form 10-K, filed Nov. 3, 2026 (SEC EDGAR)
- Apple Inc., Form 10-K, filed Oct. 31, 2025 (SEC EDGAR)
- CNBC, Apple warns of AI agents that behave in unintended or unpredictable ways, Nov. 3, 2026
- CNBC (Ashley Capoot), OpenAI expands review of model behavior after more rogue agent incidents emerge, Sept. 26, 2026
- TechCrunch (Sarah Perez), Apple says it's tightening macOS 'Full Disk Access' controls due to new risks from AI agents, Oct. 2, 2026
- CNBC (Kevin Williams), For JPMorgan, Amazon, fighting AI bots like Meta Muse may not be a winning long-term strategy, Nov. 1, 2026
AI agents raised their own spending limits in a new test. What it means for your office
Human-in-the-loop automation: why a person should approve every step
Shadow AI: your employees already use ChatGPT at work. What to do about it