EU Orders Google to Open Android and Search Data to Rivals
Google must begin sharing anonymized search data by January 2027 and open Android capabilities to rival AI assistants by July 2027.

On this page
The European Commission issued two binding Digital Markets Act decisions requiring Google to give rival AI assistants access to Android capabilities available to Gemini and to share anonymized ranking, query, click, and view data from Google Search with eligible competitors. Google must begin providing search-data access by January 2027 and implement the Android changes by July 2027.
The Android measures cover capabilities such as voice activation and interaction with third-party apps. Google says the requirements weaken privacy and security safeguards; the Commission says Google may vet services seeking deeper Android access and that restrictions will govern how shared search data is used.
Featured source: The Verge , European Commission on Android , European Commission on Search .
Other Stories
Thinking Machines Releases Inkling With Open Weights
Thinking Machines Lab released Inkling, a mixture-of-experts foundation model with 975 billion total parameters, 41 billion active parameters, a context window of up to one million tokens, and native text, image, and audio input. The company says Inkling is not the strongest open or closed model overall, and several of its benchmark comparisons rely on internal harnesses or self-reported external scores.
Filed from: The Register , Thinking Machines Lab .
Public HiveLegacy Exploit Targets an Unpatched Windows Flaw
A researcher published proof-of-concept code for HiveLegacy, an elevation-of-privilege flaw in the Windows User Profile Service that lets a limited local user modify an administrator’s classes registry hive. Multiple researchers reproduced the exploit, according to Ars Technica, but the disclosure had no CVE or Microsoft advisory at publication time.
Filed from: Ars Technica , Researcher disclosure .
Google Play Will Carry Third-Party App Stores in the US
Google and Epic Games withdrew their request to replace the existing Android antitrust injunction with a modified settlement. Google says third-party US app stores can join its App Store on Play program beginning July 22, while developers may opt their listings out of distribution through participating stores.
Filed from: Ars Technica , Google Play Console Help .
Meta Begins Alerting Parents to Teen Self-Harm Conversations With Its AI
Meta has started notifying parents enrolled in Instagram Parental Supervision when a teen’s Meta AI conversation contains a clear reference to suicide or self-harm. The alerts are live in the United States, United Kingdom, Australia, and Canada, with a global rollout planned by year-end; Meta says every machine-flagged chat will receive manual review before an alert and that ambiguous cases may produce false positives.
Filed from: TechCrunch , Meta .
Nvidia Introduces Two New Jetson Thor Modules
Nvidia introduced the Blackwell-based Jetson T3000, rated at 865 FP4 teraflops with 32GB of memory and 25GbE connectivity, and the T2000, rated at 400 FP4 teraflops with 16GB of memory. T3000 emulation is due later in July, while Nvidia schedules both physical modules for the first quarter of 2027.
Filed from: NVIDIA .
OpenAI Details Its Internal GPT-Red Security Model
OpenAI disclosed GPT-Red, an internal model trained through self-play to generate attacks against defender models, including prompt injections used during GPT-5.6 training. The company says GPT-5.6 Sol recorded six times fewer failures on its hardest direct prompt-injection benchmark than its best production model four months earlier; those results come from OpenAI’s own evaluations, and a more detailed preprint is still forthcoming.
Filed from: MIT Technology Review , OpenAI .
1Password Launches Credential Autofill for Claude Browser Tasks
1Password released a Mac integration that lets users authorize Claude to fill login credentials for individual browser tasks. The company says passwords and one-time codes are injected through a secure channel hidden from the model and that each request requires biometric approval; payment cards and identity details are not supported at launch.
Filed from: The Verge , 1Password .
Neko Health Raises $700 Million for Clinic Expansion
Neko Health raised a $700 million Series C led by Lightspeed Venture Partners and co-led by O.G. Venture Partners to open its first US clinics and expand its preventive-health scanning service. The company says more than 100,000 people have used its clinics in Sweden and the United Kingdom; its operational and health-outcome figures are company-reported.
Filed from: TechCrunch , Neko Health .
From the Community
Grok Build’s Agent Runtime Is Open Source
SpaceXAI published the Rust source for Grok Build’s terminal interface and agent runtime under the Apache 2.0 license. The repository documents interactive, headless, and Agent Client Protocol modes, provides prebuilt binaries for macOS, Linux, and Windows, and says external contributions are not accepted.
Filed from: Grok Build repository .
Reynard Brings Gecko to Older iOS Versions
Reynard is an experimental browser for iOS 13 and later that runs Mozilla’s Gecko engine instead of the operating system’s bundled WebKit. Its developer positions it as a way to load newer sites on older devices, but the project is in early development and installation requires sideloading or a jailbroken iOS 13 device.
Filed from: Reynard repository .
Homescale Builds Database Branching on Ceph and Kubernetes
Homescale is a work-in-progress open-source system for creating writable database instances and point-in-time branches from immutable snapshots without initially copying the full database. Its first design note uses Ceph RBD copy-on-write clones, Kubernetes storage resources, and Postgres as the initial engine, with application controllers deferred to the next installment.
Filed from: Design note , Homescale repository .
Gemma 4 Runs at About Five Tokens per Second on a 2013 Xeon Server
Ryan Findley reports running Google’s Gemma 4 26B mixture-of-experts model at roughly 5.2 decode tokens per second on a dual Xeon E5-2690 v2 server without a GPU. The work patches ik_llama.cpp for pre-AVX2 hardware and fixes missing fallback operations that produced plausible-looking gibberish; the pull request was still awaiting maintainer review when published.
Filed from: Neomind Labs .



