BlueFalconHD/apple_generative_model_safety_decrypted

What it solves

This project provides a method and the resulting data to decrypt and analyze the safety filters used by Apple Intelligence's generative models. It allows researchers to see the exact phrases, regular expressions, and rules that Apple uses to filter harmful content or enforce safety standards across different regions and locales.

How it works

The project uses a combination of Python scripts and LLDB (the Xcode debugger) to extract encryption keys from the GenerativeExperiencesSafetyInferenceProvider process on macOS. Once the key is obtained, a decryption script processes the encrypted safety override files located in the system library. Finally, a metadata combination script deduplicates and organizes these filters into global, regional, and locale-specific JSON files for easier review.

Who it’s for

Security researchers, AI safety analysts, and curious users who want to understand the guardrails and content filtering mechanisms implemented in Apple's on-device generative AI.

Highlights

• Key Extraction: Includes a script to programmatically retrieve encryption keys via LLDB.
• Decryption Tooling: Provides scripts to decrypt system safety overrides into readable JSON format.
• Filter Analysis: Organizes decrypted data into combined metadata files to distinguish between global and region-specific filters.
• Detailed Guardrails: Reveals specific reject, remove, and replace rules, including regex-based filtering for both user input and model output.