Researchers detail prompt injection attack bypassing Apple Intelligence protections

Security researchers have disclosed a prompt injection vulnerability in Apple Intelligence that allowed attackers to bypass the system's built-in protections and execute unauthorized commands on the on-device language model. The attack technique circumvented Apple's content restrictions, enabling researchers to force the local LLM to perform actions outside its intended safety boundaries. Apple has since corrected the vulnerability following responsible disclosure. Prompt injection attacks occur when malicious actors craft inputs that override or manipulate an AI system's instructions. The vulnerability in Apple Intelligence represents a significant concern for on-device AI systems that process sensitive user data locally. The findings highlight ongoing security challenges as companies deploy more powerful AI capabilities directly on consumer devices. Apple Intelligence, launched in late 2024, processes many queries on-device to maintain user privacy rather than sending all data to remote servers. Apple has not publicly disclosed the number of users potentially affected by the vulnerability or specific details of the security patch. The company fixed the issue prior to the researcher's public disclosure.