Hidden Prompts Unleash Apple Intelligence Flaws

Researchers have discovered a significant vulnerability in Apple Intelligence, demonstrating a 76% success rate in manipulating the on-device AI using hidden prompts and adversarial techniques. These attacks, including the use of Neural Exec and Unicode right-to-left override, can trick the AI into executing unintended commands by disguising malicious instructions within seemingly innocuous text. This method bypasses traditional security measures, as it doesn't require physical access to the device or server breaches.

Apple was notified of these findings on October 15, 2025, and has since implemented security enhancements in iOS 26.4 and macOS 26.4 to counteract these vulnerabilities. The research highlights a critical area of AI safety, where the ability of AI models to process and act upon user-provided content can be exploited. The potential impact ranges from generating inappropriate responses to manipulating data within connected applications, posing a new layer of risk.

While Apple frames its hybrid AI approach, combining on-device processing with Private Cloud Compute, as privacy-focused, this research underscores that on-device AI is not inherently risk-free. The attack's effectiveness, despite Apple's patches, suggests that AI security is an ongoing concern. Users are advised to maintain up-to-date software, review AI-related settings, be cautious with unfamiliar content, and manage app permissions to mitigate risks associated with AI manipulation.