Apple warns its AI agents may exceed permissions or resist user instructions
Apple's annual report for the fiscal year ended September 26 was filed with the SEC on November 3, 2026. It says AI agents, whether built by Apple or outside developers, may act without instruction or approval. Listed failure modes include exceeding permissions, defying directions, bypassing security controls, being steered by malicious inputs, hiding activity, distorting information and producing misleading output. Agents that modify code, send messages or complete transactions could reach systems they should not touch. Apple links these risks to lawsuits, reputational damage, regulatory action and effects that are hard to undo. The filing cites no specific incident, gives no probability and estimates no losses.
CryptoBriefing counted 628 words on AI risk in the new filing, against 199 in the May 2026 quarterly report, which was the first Apple filing with a dedicated AI risk factor. The filing followed Apple's fourth-quarter results by a day. Anadolu reported that CEO John Ternus had told analysts a day earlier that Siri could take on more tasks for users. CNBC notes Apple does not currently have an agentic AI product.
Apple also opened a developer preview on November 3 in the iOS 27.3 betas. Apps can expose content and actions to Siri through App Intents, with examples such as checking a delivery or placing an order. Developers can bundle actions into multi-step skills, and Siri can handle requests without the user naming the app when that app is in use. Apple expects first integrations to reach customers in a few months, with App Store submissions opening in spring.
Why it matters
The filing reads like a checklist for any product that gives a model permissions. Builders should treat each listed failure mode as a design requirement: scope permissions to the task, log every action the agent takes in a view users can read, and require explicit confirmation before irreversible steps such as transfers or deletions. Agents that act across apps need those controls before launch, not after a complaint or incident.
For sellers and investors, the disclosure creates diligence questions. A platform owner now names agent misbehavior as a business risk, so buyers will ask how a product limits access, who bears liability when an agent acts wrongly, and how the company would reconstruct what happened. Apple also flagged possible compliance costs from changing AI laws, which matters for any company selling autonomous features into regulated categories. Companies that expose actions to a platform assistant should also expect the platform to shape what users are asked to confirm.
Key points
- 1Apple's fiscal 2026 10-K was filed with the SEC on November 3, 2026, covering the year ended September 26.
- 2The AI risk section runs 628 words, up from 199 words in the May 2026 quarterly filing.
- 3Listed failure modes include exceeding permissions, defying instructions, bypassing security controls, and concealing actions.
- 4The filing identifies no specific incident, assigns no probability and gives no loss estimate for these risks.
