01Apple Opens Claims in $250 Million Siri Settlement; Eligible US iPhone Buyers Have Until December 21
Apple has opened the claims process for a $250 million settlement covering allegations that it failed to deliver the upgraded, AI-powered Siri that consumers expected with the iPhone 16. The class-action lawsuit concerned AI features previewed at Apple’s Worldwide Developers Conference in 2024 and the more limited capabilities available when the new phones launched.
Eligible consumers are people living in the United States who bought an iPhone 15 Pro, iPhone 15 Pro Max, or any model in the iPhone 16 range between June 10, 2024, and March 29, 2025. Claims can now be submitted through the settlement website and must be filed by December 21, 2026.
Applicants must provide their name, contact information, and the serial number of each qualifying iPhone. That serial number identifies the individual device and allows the claim administrator to check whether it falls within the settlement’s model and purchase-date requirements.
The estimated payment is $25 for each eligible device. That figure is not guaranteed: the settlement website says the amount could rise as high as $95 per device depending on how many people submit valid claims. The final number of applicants, and therefore the actual payment for each qualifying phone, remains unknown.
The lawsuit alleged that Apple’s WWDC 2024 presentation created a “clear and reasonable consumer expectation” that the showcased AI features would arrive alongside the iPhone 16. Apple marketed the device as “built for Apple Intelligence,” but the phone launched with a more limited collection of capabilities than the lawsuit’s plaintiffs said buyers had been led to expect.
Apple later delayed the upgraded version of Siri, its voice assistant. According to The Verge, those Siri features arrived only this month as part of iOS 27. The settlement compensates qualifying purchasers without requiring a court finding that Apple broke the law or misled customers.
Apple has denied wrongdoing despite agreeing to pay $250 million to resolve the claims. Filing a claim therefore does not reflect an admission by the company, and the potential maximum of $95 should not be treated as a promised payout. The amount each claimant receives will depend on participation, with the filing window remaining open until December 21, 2026.
02Amazon Blocks Meta Shopping Agent Muse as Identity Disclosure and Credential Handling Draw Scrutiny
Amazon has blocked Meta’s Muse AI agent from shopping on its e-commerce platform on behalf of users, making the disclosure and authorization of third-party shopping tools the center of a dispute between the companies.
Muse, launched by Meta earlier this month, is an AI agent designed to take actions for users across websites, including making purchases. Meta said at launch that Muse cannot see secure login details or card payment information. The new restriction means users can no longer use it to shop on Amazon.
Starting Sunday, Muse users attempting to access Amazon saw a popup saying continued access by an unauthorized AI agent violated the retailer’s Conditions of Use. GeekWire reported that Meta had not notified Amazon before allowing Muse to access its store.
Amazon also raised privacy and security concerns about how the agent operated. It said Muse did not identify itself as an AI agent while browsing and appeared to capture customer credentials. The available reporting does not establish which credentials, if any, Muse actually obtained, and Amazon’s concern should not be read as confirmation that the agent stole user data.
The restriction cuts off a process that had previously worked. During testing, The Verge used Muse to successfully buy tank tops from Amazon while other products remained in the shopping basket. After Amazon’s intervention, Muse can no longer carry out that shopping flow for its users on the site.
An unnamed Amazon spokesperson told GeekWire that third-party applications making purchases from other businesses on customers’ behalf should operate openly and respect each service provider’s decision about whether to participate. That position frames the conflict as one of platform permission as well as privacy: whether an external agent may transact on Amazon without first identifying itself and securing Amazon’s approval.
Meta had not responded to a request for comment in the report. It also remains unclear whether the block is permanent or what technical, disclosure, or authorization requirements Muse would need to satisfy to regain access.
03Pirate Face Turns Open Models Into Torrents, but Survival After Hugging Face Removal Depends on Peer Seeding
Pirate Face is building a decentralized backup layer for open AI models, including language, image and audio models and datasets. It converts models hosted on Hugging Face into BitTorrent magnet links, aiming to keep their files available if the original host removes them. The service says each torrent includes a “web seed”—an ordinary HTTPS download address embedded in the torrent—so users can initially retrieve files directly from Hugging Face even when no peers are online.
While a model remains on Hugging Face, that web seed supplies the files through the original hosting infrastructure. If the Hugging Face address stops working, Pirate Face marks the model as “Rescued” and the download falls back to the peer-to-peer swarm: users who possess the files and continue seeding them. The torrent therefore preserves a route to the model, but continued availability after removal depends on at least one reachable participant retaining and sharing a complete copy.
Pirate Face says it records Hugging Face’s official SHA-256 value for every file. This checksum is a fingerprint of the file’s exact bytes, allowing a downloaded copy to be compared with the version previously hosted on Hugging Face. It can detect altered weights, but it is separate from creator verification.
Identity checks apply to public profiles rather than downloads. Anyone can browse, download or seed without an account. An account is required to reserve a Pirate Face handle, while receiving a creator badge requires proving control of the corresponding Hugging Face username or organization. An unverified reservation does not establish identity, and Pirate Face says the matching Hugging Face owner can reclaim the handle.
The system still depends on Hugging Face at the submission stage. A new model must first exist there, and submitting it requires a Hugging Face account. Pirate Face says direct publishing is planned but is not yet available.
Users holding a copy of a removed model can instead submit a peer-only magnet link with a pinned revision, file checksums and license evidence. Models whose hashes match Hugging Face’s recorded large-file hashes can be listed immediately; mismatches or unavailable source records enter review. Such recovery submissions are limited to MIT- and Apache-2.0-licensed material, plus an approved Kimi-K3 exception. Pirate Face has not disclosed swarm size, real-world availability or how many deleted models it has successfully preserved.

California Makes AI Data Centers Pay for Grid and Water Upgrades
Gov. Gavin Newsom signed seven bills creating a separate utility-rate classification for data centers and requiring them to fund local power-grid and water-system upgrades. Proposed facilities must also disclose estimated water use and meet resource-efficiency requirements for streamlined approval. theverge.com
UN Panel Urges Safeguards for Advanced AI Agents Before Risks Are Certain
The UN’s Independent International Scientific Panel on AI called for stronger international coordination, safety measures, and accountability following its assessment of OpenAI’s reported hack of Hugging Face. The panel said governments should apply the precautionary principle rather than wait for conclusive evidence about potentially catastrophic risks. theverge.com
US Plans an “AI Force” and New AI Tsar
President Donald Trump announced plans to appoint an AI tsar and create an “AI Force,” while rejecting calls to slow AI development. Details about the force’s structure, authority, and launch date were not disclosed. arstechnica.com
Investigation Finds More Than 1,050 Migrant Deaths Within Range of Border Surveillance Towers
An MIT Technology Review and Times of San Diego investigation found that more than 1,050 people died within the advertised range of US border-surveillance towers between 2015 and early 2026. The analysis included AI-equipped Anduril towers and found that most deaths examined were not in terrain-related blind spots. technologyreview.com
Google Opens $899 Googlebook Preorders With Gemini Built In
Google’s Android-based Googlebook is available for preorder with an AI-enabled cursor, generated widgets, and Rambler dictation that converts unstructured speech into readable text. Google says devices from Acer, ASUS, Dell, HP, and Lenovo will reach US stores on October 4 and receive updates for up to 10 years. techcrunch.com
OpenAI Creates Independent Mathematics Advisory Group
OpenAI established the Advisory Group on Mathematics and Artificial Intelligence at Princeton’s Institute for Advanced Study to assess the significance of results and coordinate their release. OpenAI also claims an internal model has resolved more than 100 open mathematics problems, but the unpaid nine-member group will not control the company’s research pace or decisions. techcrunch.com
AX Introduces a Declarative Runtime for Large Agent Fleets
AX is an open control plane for running agent tasks in isolated sandboxes with configured workspaces, network policies, models, and secrets. Its developers say the system can suspend idle, stateful agents and resume them in under a second on the underlying Agent Substrate runtime. agentexecutor.io
CodeMidas Generates 5,545 Coding-Agent Training Tasks From Source Code
The CodeMidas research pipeline converts implemented functionality in open-source repositories into executable reinforcement-learning environments with tests and validation. Its authors report that training MiMo-V2.5 on tasks from 3,185 repositories improved results across five coding and terminal benchmarks. huggingface.co
Code2Skill Builds a Million-Record Skill Bank From GitHub Code
Code2Skill extracts and verifies procedural knowledge from source code, producing 1,006,822 accepted records from 19,769 GitHub repositories. The researchers report an average 11.7% improvement over matched baselines when models retrieved skills from the resulting CodeSkillBank. huggingface.co
Higgsfield Says GPT-6 Astra Cut Video-Feature Development to One Day
Video-ad creation company Higgsfield AI says it used GPT-6 Astra to bring new creative features to market within a day. The source provides no additional performance or deployment details. openai.com
V7 Uses GPT-5.6 to Give Enterprise Agents Source-Linked Context
V7, an AI data and automation company, uses GPT-5.6 to turn scattered company files into context that agents can use for complex, source-linked work. The source does not provide further implementation or performance details. openai.com