Apple opens Siri AI to every app, Anthropic open-sources FlashProve, Claude's watermark reaches older models

The Siri AI developer preview lets the new Siri AI find and use the actions and data an app exposes through App Intents. Custom actions that don't fit a system schema work too: mark an intent as a tool with @AppIntent and Siri can call it. For requests that chain several actions, you write an app skill that tells Siri how to combine them. Xcode's Siri Integration Advisor reviews your implementation, and the Siri Evaluations framework lets you script requests and measure how Siri handles them. You don't need a new entitlement or a separate program.

When your app is in the foreground or has a Live Activity running, Siri can pick its actions without the user naming the app, so voice and agents become another way into your app. Korean, Japanese and Spanish are among the launch languages, so many non-English apps can test right away. App Store submissions open in the spring, which leaves time now to clean up intents and entities and build evaluation scripts.

FlashProve lets a model provider prove that a reply came from one specific set of committed weights without revealing those weights. It's built on zkSNARKs, a cryptographic method that shows a computation was done correctly with a short proof that reveals nothing else. On a 4-bit Llama 4 Maverick (a 400B MoE), a single H100 proves about 110 tokens per secondVendor claim, and each 1,024-token request produces a proof of about 19 MBVendor claim. Checking a proof takes around 17 secondsVendor claim on a 24-core CPU, no GPU needed. The prover, a fast C++ verifier, a slower Lean-verified verifier and the Lean proofs are all released under Apache 2.0.

Right now a user can't tell whether a routing service quietly swapped in a cheaper model or an attacker tampered with the token stream. If proofs like this get practical, third-party API resellers, model evaluators and compliance checks could finally pin down which model actually answered. The cost is still 70-95x the GPU timeVendor claim of running the same requests on vLLM, so for now it fits spot checks and after-the-fact audits.

Claude-generated text carries an invisible watermark, and as of Nov 3 that covers models released before August 2, such as Claude Mythos 5, Opus 4.7 and Sonnet 4.6. It adds no tokens or hidden characters and doesn't change meaning or quality, so you don't need to touch your request or response code. On Sonnet 4.6, though, setting temperature, top_p or top_k to a non-default value can leave the watermark out. Watermark detection is in private preview for regulators, media, researchers and companies with EU AI Act obligations.

This follows the transparency rules in Article 50 of the EU AI Act, so companies building Claude into products still need to work out their own labeling duties. If you depend on the watermark, strip out sampling parameters too, including ones your framework sets for you. Detection isn't proof either: translation, summaries and short passages can lose the mark, and Claude may not be the original author of marked text.