The analysis comes from Eighth Rank, an independent AI-security research collective, and was published on 28 October; it drew wide attention only this weekend. Its subject: two background daemons — knowledgeconstructiond and intelligenceplatformd — that have run on every Mac since macOS 13 (2022). Their job, the authors say, is to build up knowledge about you: who you are, who you know, how those people are related to each other, how and how often you communicate with them, and where you go when.
The data sits in ~/Library/IntelligencePlatform: 51 SQLite databases in plaintext, guarded by TCC and Full Disk Access. The core database, graph.db, is a triple store with 181 entity classes and 278 predicates. It is fed by Contacts; faces in Photos; Calendar titles and attendees; Family Sharing including members' ages; the Medical ID emergency contact; people Siri Suggestions find in Mail and Messages; contacts donated by apps through App Intents; installed apps; location and movement history; and device signals such as screen lock, Wi-Fi and Bluetooth state. Fifty-seven of the 181 classes describe activities — sleeping, commuting, dining.
The sharpest finding concerns family. An undocumented classifier decides whether the people in your photos are family members, the researchers report, and places them in a relationship hierarchy with confidence scores. The models run locally on CoreML and XGBoost.
What that means — and what it does not. There is no indication this data leaves the device; Apple's argument has always been that exactly this kind of inference is computed locally. The criticism is not about exfiltration but about opacity and aggregation: a machine that distils kinship, closeness and daily routine into a database has produced something categorically different from a pile of files. A breach here hands over conclusions rather than records, the researchers argue — including whom you would call in an emergency.
The authors released a tool called iKnow so users can inspect their own graph locally, and a second piece walks through what third-party access could do with it. The Hacker News discussion, where the post reached several hundred points over the weekend, mixed alarm with scepticism — not least about the plainly AI-assisted prose. The checkable artefacts survive that objection: paths, schemas and daemons can be inspected on your own machine.
Apple has, as far as is publicly known, never documented the classifier. Anyone who wants to know what their Mac infers about their family currently has to go and look.




