Security Flaw Discovered in Apple Intelligence on iPhones

A serious security flaw has been discovered in Apple Intelligence on iPhones. Recent research indicates it is vulnerable to hidden prompt attacks.

Researchers at RSAC Research found they could manipulate Apple Intelligence with a 76% success rate in 100 tests using hidden prompts, adversarial prompts, and Unicode tricks, demanding a reassessment of existing privacy assurances.

These attacks can be initiated through carefully crafted text that tricks the AI into performing specific tasks contrary to the user's intent, without requiring the user's iPhone to be stolen or their password to be unlocked. This could be exploited when AI is used to read, summarize, modify, or act on behalf of apps.

The RSAC researchers tested the device-based large language models embedded in Apple's operating system. This is significant because third-party apps can access Apple Intelligence through system tools and APIs.

Two main techniques were employed in the attacks. One involved using unconventional prompt formats designed to confuse the model and elicit specific responses, while the other leveraged Unicode's right-to-left override functionality.

After researchers disclosed the findings to Apple on October 15, 2025, Apple strengthened its protections through updates in iOS 17.5 and macOS 14.5.