A researcher published findings today showing that every Mac since macOS 13 Ventura — released in 2022 — has been silently running two background daemons that build a Mac hidden knowledge graph about you, your family members, your contacts, and your behavioral patterns. Apple has never publicly documented any of it. To prove the point, they released an open-source tool called iKnow that lets you read exactly what your Mac has already inferred about you. The Hacker News discussion hit 342 points and 194 comments within hours of posting.
What Your Mac Has Already Decided About Your Family
The database lives at ~/Library/IntelligencePlatform and is built by two daemons Apple describes as software that will “analyze content on this device to help build a knowledge graph representing the device and the user.” The core file, graph.db, uses a triple-store model with 181 entity classes and 278 predicates. One developer who audited the full database set found more than 350,000 records across 51 SQLite files, each relationship tagged with a confidence score between 0 and 1.
The graph ingests data from: Contacts, Photos (including facial recognition and face clustering), Calendar, Family Sharing (including the recorded age of every family member), the Medical ID emergency contact in Health, events and contacts from Mail and Messages surfaced by Siri Suggestions, location history, installed apps, and — critically for developers — contacts donated through App Intents. Machine learning runs on top: XGBoost classifiers work from 144 behavioral features to infer family relationships, while embedding models resolve when two separate records represent the same person. Moreover, the system uses Wikidata QIDs as its topic ontology, meaning your behavior is mapped against a global knowledge graph of entities. The result is that your Mac can infer household composition, child ages, and communication networks without you ever explicitly providing that information.
Developers: Your App Intents Are Feeding the Mac Knowledge Graph
If you have built a macOS app that implements App Intents and donates person or contact information, that data flows into the IntelligencePlatform graph. Apple’s App Intents documentation does not make this connection explicit. Furthermore, the graph has been building since macOS 13 — well before Apple Intelligence launched as a branded product in 2024 — meaning developers who adopted App Intents early have been contributing user data to this system for years. If you are shipping a contacts-aware macOS app, it is worth auditing what you donate through App Intents and understanding where that data goes downstream.
The Real Risk: AI Agents Can Read the Mac Knowledge Graph
Apple’s standard privacy response — “it stays on your device” — is accurate as far as it goes. However, on-device does not mean inaccessible. The IntelligencePlatform databases are plaintext SQLite files, readable by any process with Full Disk Access. AI coding agents — Claude Code, Codex, Cursor, Gemini CLI — routinely request that permission. When a coding agent gets Full Disk Access, it can read Apple’s inferences about your family relationships, your communication patterns, and your location history. This is not theoretical: in October, Apple announced plans to tighten Full Disk Access restrictions after Meta’s Muse AI agent accessed a journalist’s private iMessages. The IntelligencePlatform exposure almost certainly contributed to that decision.
What You Can Do Right Now
To inspect what your Mac has built, run the iKnow tool — a Python standard-library web dashboard that opens locally in your browser and reads the graph in read-only mode. It requires Full Disk Access for the terminal, works on macOS 26.6.2 through 27.0.1, and nothing leaves your device. If you want the daemons to stop building the graph, the supported path is System Settings → Apple Intelligence & Siri and disabling the feature. Deleting graph.db directly just triggers an immediate rebuild. Additionally, if you are granting AI coding agents Full Disk Access, now is a good time to reconsider the scope of that permission — specifically whether blanket file system access is actually required for your workflow.
Apple’s on-device AI story has always depended on a tacit assumption: local means private. The full anatomy of this knowledge graph makes that assumption harder to hold. A 350,000-record behavioral profile with family relationship inference and Wikidata-mapped interests is not a small privacy footprint — it just happens to sit on your hard drive instead of a server. In the AI agent era, with tools that freely request Full Disk Access, the distinction matters considerably less than it used to.
