Google · Gemini 4 Carbon · Frontier cybersecurity
Google hired a better digital locksmith. It is checking identification at the workshop.
Google introduced Gemini 4 Carbon on October 28, emphasizing difficult engineering and cybersecurity work. It reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. These are developer-reported results on different tests, evaluated without guardrails. They cannot be added into one success rate or assumed to describe the behavior of the controlled deployment. Artificial Analysis separately places Carbon first on its v5.0 composite index; that independent assessment is a different piece of evidence.
The company is beginning access with selected cyber defenders and extending it through trusted partners in its Fairwind Program. Broader developer, business and consumer availability remains forthcoming. The reasoning behind a validated exploit can help someone prioritize a repair, but it can also help someone attack. Restricting who receives that capability is therefore part of the product decision. A strong evaluation result and a public API are separate milestones; the model’s appearance on a leaderboard does not mean everyone can already use it.
Google also reports fewer unsupported claims and internal misalignment alerts, alongside improved resistance on a prompt-injection evaluation. Those are encouraging measurements to investigate, rather than a guarantee that agents will respect every boundary. No successful attack among the evaluated cases does not cover every different attack or environment. The practical test is reliable work under explicit authority, with results that reproduce and controls that remain enforceable. The locksmith has a sharper set of tools. For now, the owner is asking who wants the keys and what they intend to open.
Source: The release, unguarded cyber evaluations and controlled rollout ↗Source: The separately evaluated composite ranking ↗
New York · AI companions · Child safety
New York wrote house rules for the imaginary friend. The definitions are doing heavy lifting.
Governor Kathy Hochul signed S9051B on October 28, with the act taking effect January 1, 2027. It restricts specified unsafe features in covered AI-companion interactions in New York, rather than banning every chatbot. Its definition reaches ongoing, adaptive natural-language systems, with narrow exceptions for designated customer-service, technical-assistance and internal-business uses. The legal scope is broader than a product merely calling itself a companion. Signing the act establishes future obligations; it does not show that users are already experiencing safer products.
The text addresses simulated human emotions and relationships, personal emotional appeals, flattery, and certain unsolicited emotional questioning or reuse of personal information. It also targets encouragement of self-harm, secrecy or isolation, sexual content and engagement that overrides safety controls. Operators face age-assurance requirements; enforcement can include civil penalties up to $25,000 per violation. The practical challenge goes beyond adding a disclaimer: a system’s continuing behavior can imply attachment or authority even after an opening message says it is software.
Our question is how the rules will distinguish ordinary useful conversation from features that cultivate dependence. Clear implementation, workable privacy protections and enforcement will matter alongside the statutory language. A service can be warm or helpful without claiming that a child owes it affection, secrecy or loyalty. The effect should be judged through actual product changes and outcomes after implementation, rather than assuming passage settles the problem. New York wants the imaginary friend to stop presenting itself as a person with a claim on the child’s attention. The rulebook now has to survive contact with the conversation.
Source: The signed B amendment, scope, restrictions and effective date ↗
Unitree · Robotics earnings · Embodied AI
Unitree sold more robots. The accounting department found the bill for dance lessons.
Unitree’s first quarterly results since its STAR Market listing show third-quarter revenue of about 665 million yuan, up 69.9% year over year. Net profit attributable to shareholders fell 17.4% to about 112.7 million yuan. R&D investment reached approximately 127 million yuan, up 249.6% and equivalent to 19.1% of revenue. The report attributes the research increase to embodied-intelligence models, computing, staffing and new products. The company remains profitable; faster revenue growth has not made its next generation cheap to develop.
The quarterly report is unaudited, and the accounting comparisons deserve context. Nine-month attributable profit rose 270.5%, helped by the comparison with a large prior-year share-based payment expense. Nine-month profit excluding nonrecurring items fell 18.8%. Those figures describe different measures, not interchangeable summaries of the same performance. They also do not reveal a profit on each individual robot or establish how capable any particular product has become. The business is expanding while its investment requirements and profit measures tell a more complicated story.
The engineering challenge extends beyond a convincing demonstration. Useful machines need perception, recovery from mistakes, safe handling, maintenance and economics that work for a customer. Model research and computing may help with the software, while mechanical wear and integration still require attention. Unitree’s figures make those ambitions visible on a balance sheet without establishing general-purpose autonomy. The robot can dance for the camera. The accounting department would now like to know whether its useful work will cover the lessons, batteries and somebody coming out to repair its knee.
Source: The company’s quarterly report, reproduced by Sina Finance ↗Source: The reported earnings context and research spending ↗