EU Orders Google to Open Android to Rival AI Despite Its Safety Warnings

01Google Says Opening Android to Rival AI Will Endanger Users. The EU Just Ordered It Anyway.

The European Union told Google on Thursday to give competing AI assistants and search engines deeper access to Android and Google Search, two rulings that pry open the platforms Google has kept for itself. The decisions stem from the bloc's Digital Markets Act. Google's answer is not a legal counterargument but a warning about its own users: the company says the changes could endanger their privacy and security.

That sets up a direct fight over a single question. Who controls the door between a phone and the AI behind it.

For rival assistants, the door has been locked. Google preloads its own AI on Android and routes queries through Google Search by default, giving competitors no clean way to sit where users actually tap. The EU orders greater access to key parts of both. A rival assistant could reach the entry points on Android that Google reserves, and rival search engines gain access to the search data Google has kept proprietary, according to the two decisions.

Search data is the harder concession. Training and ranking a competitive engine requires the click and query signals that only an incumbent at Google's scale accumulates. Forcing Google to share that data attacks the moat directly, not just the shopfront. The Verge reported the rulings could weaken Google's control over two of the tech industry's most important platforms.

Google's public position reframes access as exposure. Opening Android's internals and handing search data to third parties, the company says, creates paths for user information to leak and for security to weaken. The argument turns every interoperability requirement into a potential breach. It also happens to describe exactly the mechanism that would let a competitor build a rival to Google's own products.

Neither side disputes the underlying mechanics. More access for rivals means more parties touching Android's plumbing and Google's data. The EU treats that as competition working as intended. Google treats it as risk that lands on users. Both can point at the same order and call it proof.

The rulings were handed down Thursday, and the practical fight now moves to how Google implements them. Every technical detail of how access is granted becomes a place to argue that a safeguard is really a lock.

02Your password manager will now hand Claude the keys to your logins

1Password released a browser integration that lets Anthropic's Claude reach stored credentials, including usernames and passwords. The company says the feature, called 1Password for Claude, lets users authorize the chatbot to finish multi-step jobs such as booking travel or managing online accounts. Claude enters the logins itself. The human stops typing them.

Google is moving the same direction. It began letting people connect frequently used apps to AI Mode in Search, so the assistant can link to and operate those services directly. Google says the connections stay secure and under user control, and it frames the change as convenience: your go-to services, reachable inside a search query.

Strip the product language and both do one thing. The assistant layer is no longer just reading your information. It now holds credentials and account access to services where you are already signed in.

That resets the risk math. An assistant that could only read your data leaked information if it was tricked. An assistant that holds your password vault and logged-in sessions can act: submit forms, move money, change account settings on your behalf. The blast radius of a manipulated instruction grows from what the model can see to what your accounts can do.

The two grants differ in kind. 1Password exposes a credential store, the master list that unlocks everything behind it. Google's Search links reach specific connected apps and the actions they permit. Both shrink the same gap, the one between an assistant reading about your accounts and an assistant using them.

For developers wiring these integrations, the design question becomes scope: which credentials and which actions an agent can touch once authorized. For everyday users, the decision moves earlier, to the moment of granting access. The practical check before clicking approve is which already-logged-in services the assistant can then reach, because that list defines what a bad instruction could spend or expose.

Neither company has published limits on which account actions an authorized assistant may take without a second confirmation. That threshold, more than the feature launch, will determine how much a single misdirected prompt can cost.

03A 0.8B open parser reads full document pages, while big vision models still stumble on real layouts

Most document pipelines chain separate models: one to find text, another for tables, a third for formulas. OvisOCR2 takes the whole page at once. Given a document image, it outputs a Markdown version in natural reading order, covering text, formulas, tables, and visual regions in a single pass. The model holds 0.8 billion parameters, small enough to run outside a datacenter.

Getting there took a data engine built for the problem. Its authors say they mixed filtered annotations from real documents with synthetic pages, rendering each synthetic image and its Markdown target from the same HTML source so the label always matches the pixels. Training ran supervised fine-tuning first, then reinforcement learning. The pitch is a parser that reads a page the way a person does, top to bottom, and hands back clean structure.

The reason that job stays hard shows up in a second paper released alongside it. Vision language models already post strong numbers on DocVQA, ChartQA, and MMLongBench-Doc. Those scores say little about real files, where length, layout density, modality, and question difficulty pile up together and hide the actual point of failure.

SynthDocBench tries to pull those factors apart. It is fully synthetic and controls each variable on its own, so a wrong answer can be traced to long context, or crowded layout, or a specific mix of chart and prose, rather than blamed on the document as a whole. Its authors frame it as a diagnostic for long-context visual document understanding, and the diagnosis is unflattering: capable general-purpose models keep tripping when a page combines many elements at once.

For anyone wiring up an OCR or document ingestion stack, the two papers read as one decision. A specialized 0.8B parser aims squarely at conversion, page to Markdown, and nothing else. A frontier VLM promises to answer questions about the same page but, on controlled real-world layouts, still misreads them. The benchmark gives builders a way to measure that gap before it reaches production, one factor at a time.