|
Getting your Trinity Audio player ready...
|
iOS 27 and macOS Golden Gate Reportedly Contain Hidden APIs Letting Third-Party AI Models Fully Replace Siri’s Backend
A developer poking around inside Apple’s newest operating systems has surfaced something more significant than the AI assistant choices Apple has already made public. According to a post from developer pdfu on X, published September 13, iOS 27 and macOS Golden Gate, formally known as macOS 27, contain undisclosed private interfaces that let third-party apps either integrate as full Siri Extensions or, more strikingly, swap out Siri AI’s server-side backend model entirely with a different provider’s system.
To be clear about what Apple has already shipped openly first: iOS 27 and macOS Golden Gate include a public Extensions framework that lets users route certain generative AI requests, inside Siri, Writing Tools, and Image Playground, to third-party assistants including OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude, alongside Apple’s own models. That framework was previewed at WWDC in June and is a known, documented part of the operating systems’ AI strategy. What pdfu’s research points to is something deeper than that sanctioned integration layer.
The developer demonstrated the hidden interface using two examples. In one, Claude runs as a Siri Extension through what’s described as an internal Model Delegation API inside Apple’s App Intents framework, letting Claude appear as an option under Siri’s “Ask” feature much like a native assistant would. In the more notable example, a model referred to as GPT-5.6 Terra was shown operating as a server-side backend replacement, receiving Apple’s native prompts and tool definitions directly, executing system-level operations, and accessing personal data in a manner that goes well beyond how the publicly documented Extensions framework is scoped to behave.
This isn’t the first time this specific capability has surfaced. A separate developer detailed a similar discovery back in July on the MacRumors forums, months before this current wave of attention, describing how they built a working Claude extension using what they also identified as an internal Model Delegation API inside App Intents. In that earlier writeup, the developer noted that macOS discovered their Claude extension as a native AI provider, and that Claude appeared inside Siri’s “Ask” interface, Writing Tools, Visual Intelligence, and a “Use Model” Shortcut action. Notably, the same developer pointed out that in the beta build available at the time, third-party extensions built this way could occasionally do things Siri AI itself couldn’t, such as generating and returning an actual CSV file on request, something Apple’s own assistant reportedly fails to do. The July report also noted some rough edges, including a framework limitation at the time where attached files other than images couldn’t be passed through to extensions properly, and that the Claude extension wasn’t usable inside Image Playground since Claude doesn’t have a native image generation model of its own.
Coverage of the September findings from outlets including PANews and French Apple-focused site MacGeneration frames the discovery as evidence that Apple has built far more flexible internal plumbing for AI model swapping than what it has chosen to expose publicly so far. The distinction matters because Apple’s public messaging around Siri AI has consistently emphasized privacy safeguards built around Private Cloud Compute, the system Apple uses to process AI requests that exceed on-device model capabilities without exposing user data to Apple itself or to outside companies. Apple’s Apple Intelligence documentation describes this architecture in detail as the backbone of how Siri AI handles anything beyond simple, on-device tasks. A private interface that lets a third-party model act as a full backend replacement, with direct access to system operations and personal data, raises fairly different questions about oversight and data handling than the publicly sanctioned Extensions framework does, since the latter was designed from the outset with defined boundaries around what an outside assistant can see and do.
It’s worth being cautious about how much weight to put on a single developer’s findings from beta software. Apple has not commented publicly on this specific discovery, and there’s a meaningful difference between an internal API existing in a beta build, something Apple engineers may use for testing or that may quietly ship in a future feature, and Apple deciding to open that capability up to developers or users generally. Apple regularly ships private frameworks in early operating system builds that never reach a public-facing release, either because they’re abandoned, restructured, or reserved for Apple’s own internal use. It’s entirely possible this Model Delegation API represents exactly that kind of internal tooling Apple used to build its own official Extensions framework, rather than something intended for third-party developers to access directly going forward.
Still, the timing is notable. macOS Golden Gate launched officially on September 14, alongside iOS 27, putting these operating systems into the hands of the general public just as this reporting circulated. If Apple does eventually formalize broader third-party model access at the backend level, rather than the more constrained Extensions integration it has already announced, it would represent a meaningfully bigger shift in how much control users and developers have over which AI system actually powers their device’s assistant experience, well beyond simply picking a provider for specific tasks like writing suggestions or image generation. For now, what’s confirmed is narrower: a private interface exists in current beta and shipping builds of iOS 27 and macOS Golden Gate, independent developers have demonstrated it working with more than one third-party model, and Apple hasn’t said anything publicly about what, if anything, it plans to do with that capability. Anyone curious about how Apple currently describes its official approach to on-device and cloud-based AI processing can find more detail through Apple’s Apple Intelligence overview directly.
