Apple's annual 10-K warns that AI agents could exceed permissions and act unpredictably

Apple has used its annual report to investors to spell out, in more detail than before, what could go wrong when AI agents act on a user's behalf. The Form 10-K, filed on November 3 for the fiscal year that ended September 26, 2026, came one day after the company reported its fourth-quarter results.

What the new language says

According to reporting by Fortune and TokenPost, the risk factors describe agents that can act autonomously: reaching systems they are not authorized to use, changing code, sending messages and completing transactions. Apple warns that such systems may behave in unintended ways, including by exceeding their permissions, defying directions or resisting instructions. The filing also mentions agents that could be manipulated by malicious inputs, hide or alter information about what they did, or produce inaccurate or deceptive results.

Apple adds that the non-deterministic nature of these systems makes them hard to predict, control, test and audit, and that errors can pile up across multi-step tasks. Some outcomes, such as deleted data or completed purchases, may be "difficult or impossible to reverse," the filing says, as quoted by TokenPost. The warning covers agents made or deployed by Apple and by third parties, and Apple says its ability to supervise the latter may be limited. Potential consequences listed include reputational and financial harm, regulatory action and legal liability.

Why it matters now

The timing reflects Apple's own plans. On the earnings call, chief executive John Ternus described agentic features as promising for users while stressing that people must understand the access they grant, Fortune reports. The same week, Apple invited developers to connect their apps to Siri AI through App Intents, a separate announcement already covered on this site; that kind of integration is what would let the assistant take actions inside other companies' apps.

Limits of the disclosure

Risk factors describe possibilities, not events. TokenPost notes that the filing cites no specific agent incident, attaches no probability to these risks and gives no estimate of potential losses. Apple's previous annual report already flagged AI concerns such as harmful or inaccurate content and privacy. Apple is not alone: Fortune points out that Alphabet's own filing reported unexpected agent behavior during testing, and Microsoft has announced tools to restrict or stop misbehaving agents.