Security researchers have documented a hidden system that ships on every Mac: a pair of background processes that quietly build a detailed, queryable knowledge graph about the person using the machine and everyone around them — their contacts, their family relationships, their daily routines, even their face data — with almost no official documentation from Apple.
The analysis, published October 28 on the Eighth Rank blog by researchers Dario Pasquini and Michał Bazyli, resurfaced in a major way this week when it hit the front page of Hacker News on Friday, drawing more than 750 points and a lengthy debate about on-device privacy. Kabeh sing ditulis diklumpukake ing macOS 26, versi 26.6.2, miturut penulis. For more context on this story, see our ongoing latest AI developments.
Loro Daemon, Seket Siji Database
The system is built by two daemons that have shipped with every Mac since macOS 13 arrived in 2022: knowledgeconstructiond and intelligenceplatformd. Apple's own documentation for the mechanism amounts to a single man page, which describes knowledgeconstructiond as a daemon that "analyzes content on this device to help build a general purpose knowledge graph" — and little more.
According to the researchers, the daemons cross-reference nearly everything the operating system can see: Contacts, Calendar, Photos (including face data), Messages and Mail, location, Screen Time, Family Sharing data — including the age of every member of a household — the Medical ID emergency contact in Health, events and contacts surfaced by Siri Suggestions, people donated by third-party apps through App Intents, and the list of installed apps.
The result lives in ~/Library/IntelligencePlatform: fifty-one SQLite databases stored in plaintext, protected from casual access by macOS's TCC Full Disk Access gate. Inti sistem kasebut yaiku graph.db, telung toko sing ditulis ing kosakata 181 kelas entitas lan 278 predikat. Every row encodes a subject, a predicate and an object — a person, a relationship type, and another person, place, organization or event — with each edge carrying confidence scores, source counts and timestamps.
Uripmu, Digolongake
Jembare ontologi yaiku ing ngendi riset kasebut bisa ditemokake. The researchers catalogued all 181 entity classes and found that 57 of them — nearly a third of the vocabulary — describe activities: sleeping, exercising, mindfulness sessions, going to bed, commuting, dining. Alongside the graph sit auxiliary tables recording who appears with whom in your photos and a communication log capturing who messaged whom, when, in which app, and in which direction — metadata only, without message content.
Forty-six of the predicates are relationship types that reify the edges between people, which is how the system captures not just who you know but how they relate to each other. Kapabilitas kasebut nyebabake pangguna Hacker News ngringkes panemuan kasebut kanthi blaka: saben Mac ngirim klasifikasi anggota kulawarga sing ora duwe dokumen.
Grafik dudu metadata nganggur. The researchers describe it as a queryable store that large language models retrieve from — the substrate that lets Apple Intelligence answer questions about "my sister's photos from last summer" without sending data to the cloud. Apple's own researchers published a paper on the underlying graph-construction techniques in 2023, but the researchers note that almost nothing has been published about how the system is integrated into the shipping operating system.
Open-Sourcing Peek - lan Peringatan
To let users inspect their own graphs, the researchers released iKnow, an open-source dashboard that reads the contents of ~/Library/IntelligencePlatform and renders what the system believes about you and your family.
Bagian pungkasan kertas kasebut sengaja provokatif. The authors run a thought experiment — which they note applies to any personalized system — showing what happens if such a knowledge base ends up in the wrong hands, combining it with an "abliterated" language model (one with safety training removed) to demonstrate how a "privacy single point of failure" can be weaponized against a user in ways that individual data leaks cannot.
Gambar sing luwih gedhe
Panemuan kasebut kanthi kikuk nglawan marketing privasi-pisanan sing wis nemtokake strategi AI Apple: pangolahan ing piranti sejatine luwih pribadi tinimbang alternatif maya, nanging riset iki nuduhake yen "tetep ing piranti sampeyan" ora ateges "tetep cilik." Sistem lokal sing ngerti jadwal turu, lungo, wit kulawarga lan riwayat foto sampeyan minangka target sing dhuwur - apa ancaman kasebut minangka aktor ala sing nduweni akses singkat menyang laptop, kerentanan piranti lunak ing mangsa ngarep, utawa agen lokal sing bisa banget dijaluk kanggo nyalahi panggunaan apa sing bisa ditakoni.
Ora ana data, peneliti nandheske, ditularake saka piranti dening daemon dhewe; analisis babagan apa sing disimpen sacara lokal lan carane. Nanging amarga asisten AI entuk akses sing luwih jero menyang konteks lokal, titik inti peneliti: pangguna ora diwenehi gambaran sing jelas babagan apa sing digawe mesin dhewe, lan kudu ngowahi jawaban kasebut dhewe.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
