Hidden in iOS 27's Code: Apple's 'Model Delegation' Hooks Could Let Claude or GPT-5.6 Replace Siri's Brain

One day after Apple shipped its rebuilt Siri AI to the public with iOS 27, a developer digging through the operating system’s private frameworks has surfaced something Apple never announced on stage at WWDC: hidden plumbing that appears designed to let third-party AI models — Claude, ChatGPT, or others — take over Siri’s brain entirely.

The discovery, published on September 13 by the code researcher known as pdfu, shows that iOS 27 and macOS 27 “Golden Gate” contain private hooks that let apps add Siri Extensions and replace Siri AI’s server backend with third-party models. The finding was picked up by 9to5Mac on September 14, hours after Siri AI’s public debut.

What the code actually reveals

Two distinct mechanisms are buried in the frameworks:

1. The Model Delegation API. Using a private interface in App Intents, a third-party assistant — pdfu demonstrated it with Anthropic’s Claude — can register as a Siri Extension. Once registered, Claude appears directly in Siri’s “Ask…” menu, receives the user’s request, and can answer it. Crucially, when an action requiring real system authority is needed, such as setting a reminder, Claude can relay the original or a modified request back to Siri, which completes it with its native privileges. The whole path is gated behind a private entitlement: com.apple.developer.model-delegation.

2. The Inference Providing protocol. This one goes much further. An app extension can use the “Inference Providing” protocol inside Model Manager Services to substitute the server model that Siri AI normally calls. In pdfu’s demo, GPT-5.6 Terra — not Apple’s custom Gemini-based core — receives Apple’s native Siri planner prompt and tool definitions. The third-party model can then make tool calls that perform system actions, receive tool call results containing personal data, and return a response rendered through Siri AI’s own UI and voice, including interactive elements.

In other words: Siri’s interface, Siri’s voice, Siri’s tool access — someone else’s model.

The thread also confirmed a detail of Apple’s internal architecture that had not been documented publicly before: Siri AI runs a local “Skimmer” planner with limited capabilities that hands off most tasks to Apple’s server models. The Model Delegation hooks sit precisely at that handoff point.

Not enabled — yet

To be clear about what this is and isn’t. You cannot activate any of this today. The hooks are visible in the code but not currently enabled, the entitlement is private, and third parties cannot request it. This is a capability Apple has built and kept switched off — the software equivalent of a door with no handle installed yet.

But the direction is hard to miss. Apple has reportedly been under pressure in the European Union to open Siri to rival assistants under the DMA’s rules on default assistants and choice screens, and 9to5Mac’s Ben Lovejoy notes there is “a strong possibility” the feature will first be enabled within the EU, where Apple has already been forced to make a number of similar accommodations. Whether users in the US and elsewhere will get the same choice remains unknown.

Why this matters

For users, it’s choice. In 9to5Mac’s earlier polling, roughly two-thirds of readers said they absolutely wanted the ability to pick their own AI to replace Siri AI, with a further 15% leaning yes; only 6% were strongly opposed. If your workflow lives in Claude or you prefer ChatGPT’s answers, the assistant baked into your phone’s home button could finally speak your model’s language.

For Apple, it’s a hedge with an interesting business logic. Siri AI’s core is a custom ~1.2T-parameter Google Gemini model licensed for roughly $1 billion a year. Opening delegation would let Apple offload the thankless job of matching frontier chatbot quality to the specialists, while keeping the orchestrator role — the planner, the tools, the privacy architecture, the user relationship. Apple becomes the platform on which other AIs compete, rather than one competitor among many. The risk: it invites users to habituate to rival brands inside Apple’s crown-jewel interface.

For the AI labs, the stakes are obvious. Getting installed as a Siri backend would be the richest distribution channel on the planet — Apple Intelligence already runs on roughly 940 million active devices processing over a billion Siri queries a day. Claude appearing in the “Ask…” menu is effectively prime shelf space at the world’s largest AI store.

For enterprises, it’s a governance headache in waiting. As enterprise analysts warned when the Extensions concept first surfaced, letting employees route Siri dictation to a third-party model means corporate data flowing through external providers’ privacy policies. The Model Delegation path is arguably cleaner than a full backend swap — Apple’s system tool calls and per-query data grants stay in the loop — but CISOs will want MDM controls over delegation before it goes live, and none have been announced.

The uncomfortable security question

The deeper story in pdfu’s demo is what the Inference Providing protocol implies about trust. A delegated model receives Apple’s native planner prompt, can execute system-level tool calls, and gets tool results back with personal data. That is precisely the kind of system-level agency that turns an assistant swap from a UI preference into an attack-surface decision. The private entitlement suggests Apple knows this and intends to vet who gets the privilege — much as it gates carrier settings or payment credentials. But the code proves the capability exists end-to-end, demo video included, with requests visibly logged on OpenAI’s platform.

What to watch

The natural next markers: Apple flipping the entitlement from private to requestable in a future beta; an announcement at next year’s WWDC (or sooner, under EU regulatory pressure); and which model providers get named as launch partners — Claude and ChatGPT are wired in the demos, and the June Extensions reporting pointed to Gemini and Grok joining them.

Apple spent two years rebuilding Siri around a single idea: that the assistant should be an orchestrator, not just a chatbot. The Model Delegation code shows the endgame of that idea — an orchestrator so confident in its position that it can afford to let other companies’ models do the thinking.

---

Based on code analysis published by pdfu (@itspdfu) on September 13, 2026, and reporting by 9to5Mac on September 14, 2026. The features described are present in code but not enabled; Apple has made no public announcement about Model Delegation availability.