IT News Review by Control F5 Software: Humanoid robots create a new category of work – people who teach them how to work

Humanoid robots create a new category of work: people who teach them how to work

The development of humanoid robots has led to the emergence of a new type of activity: contractors who film themselves doing household and everyday tasks to generate training data. Essentially, for robots to learn how to perform activities from the real world, companies need large volumes of visual and behavioral data captured from a human perspective.

This stage shows that the problem is no longer just the construction of the robot, but also the accumulation of the right data for safe and useful behavior. Just as language models needed large bodies of text, humanoid robots need large bodies of human action. This shifts the focus from spectacular demonstrations to the invisible infrastructure that makes learning possible.

For the industry, it is a clear sign that embodied AI is already starting to generate new economic chains around data collection, training, and operationalization in the physical world. In other words, advanced robotics is no longer just about hardware and algorithms, but about the entire ecosystem needed to transform movement and decision-making into repeatable processes.

Copilot and the trust limit in AI. Is the product just for entertainment?

The terms of use for Copilot included the wording that the product is "for entertainment purposes only" and that users should not rely on it for important advice. The reaction arose at a time when the company is heavily promoting Copilot to enterprise clients, and the contrast between the commercial message and the legal text quickly drew attention. Microsoft later conveyed that the wording is an old one and that it will be updated to better reflect the current way the product is used.

The subject is relevant for the software market because it once again shows the distance between the product promise and the actual level of legal responsibility assumed by AI providers. For companies, the practical message is clear: generative tools can accelerate work, but they do not eliminate the need for validation, governance, and human control in critical flows.

Claude Code becomes more expensive for use in third-party tools

Anthropic announced that Claude Code subscribers will no longer be able to consume the limits included in the subscription when using the product through third-party tools, including OpenClaw. Instead of this model, the company is moving to a separate, pay-as-you-go system for this type of use. The change was communicated via email to clients and took effect on April 4.

For the development ecosystem, the move signals a maturation of business models around coding agencies. As more teams integrate such products into their own interfaces, IDEs, and automated flows, the cost of infrastructure and the delineation between the core product and extended integration become increasingly important.

ElevenLabs enters the music generation space

ElevenLabs launched the ElevenMusic app on iOS, allowing users to generate music with AI and discover tracks created by other users. The move expands the company's positioning beyond voice models and puts it in more direct competition with platforms like Suno and Udio. The app had been listed in the App Store a few weeks before the official launch on April 1.

From a business perspective, the launch shows how specialized model providers are looking to climb the value chain and build end products, not just infrastructure. For the software market, this type of expansion matters because it reduces dependence on a single category of model and opens new sources of monetization in the generative media space.

Google adds prompt control in Vids

Google introduced new features in the Vids app for controlling avatars through text prompts, support for Veo 3.1, direct export to YouTube, and recording via a Chrome extension. Users can ask avatars to interact with products, props, or equipment and can modify the appearance of characters, clothing, and backgrounds through natural language. Google claims that the app maintains character consistency even in more dynamic scenes.

The relevance for the enterprise space is evident: internal and commercial video is increasingly becoming an AI-assisted software flow, where editing and generation approach the logic of a low-code tool. For product and marketing teams, the direction indicates a stronger automation of video content without leaving the already used collaboration suites.

Sonder bets on an intentionally difficult onboarding

The Sonder app tries to break out of the dating platform mold through a deliberately more tedious sign-up process, designed to filter superficial profiles and attract more engaged users. The founders say that the starting point was frustration with the repetitive and substance-lacking experience of existing apps.

From a digital product perspective, the case is interesting because it goes against the classic reflex of reducing any friction. In certain categories, stricter onboarding can become a quality mechanism, not a conversion problem, especially when the goal is relationship density and trust in the community.

Cognichip wants to use AI for chip design

The startup Cognichip has raised $60 million to develop a deep learning model to assist engineers in designing new chips. The company addresses a structural problem in the industry: semiconductor design is slow, very expensive, and can take years before mass production. According to TechCrunch, just the design phase can take up to two years before the physical layout.

For the AI ecosystem, the theme is strategic. If models can accelerate not only software but also the hardware design that powers AI, the impact shifts directly into the underlying infrastructure of the industry. Here we are no longer talking just about productivity, but about time to market, development cost, and the ability to iterate faster on specialized hardware.

Slack receives a massive AI-centered upgrade

Salesforce has unveiled an updated version of Slack, with 30 new AI-based features. The most important change concerns Slackbot, which takes on a role closer to an agent: it can draft emails, schedule meetings, and search for specific information in the inbox, continuing the open direction set by the January update.

For companies, the move confirms the transformation of collaboration platforms into operational interfaces for software agents. Value no longer comes just from messaging, but from the ability to connect conversation with task execution, data, and processes. This is an important direction for organizations that are building productivity around their internal digital suite.

Google allows Gmail address changes in the US

Google has started allowing users in the US to change their Gmail address without losing access to existing data. The feature is activated from the account settings, and users can modify their username once every 12 months. The old address remains preserved as an alternative address, and authentication in Google services can continue with both options.

It is a small change on the surface, but important in the logic of digital identity. For users and companies, the email address is a persistent element in authentication flows, reputation, and communication. The ability to update it without a complete migration reduces administrative friction but also introduces new implications for support, security, and compatibility between services.

More and more Americans accept the idea of an AI boss

A Quinnipiac poll shows that 15% of Americans would accept having an AI program as a direct supervisor, responsible for task allocation and scheduling. The study was conducted on 1,397 adults in the US, from March 19-23, 2026, and included broader questions about AI adoption, trust, and workplace fears.

Even though the majority do not support this scenario, the mere fact that the percentage exists is relevant for work management software. The direction suggests that the idea of automated activity coordination is becoming less theoretical and more of a product concern, especially in organizations that already operate based on tasks, scheduling, and digitized metrics.

AI adoption is increasing, trust remains low

A result from Quinnipiac polls shows that the use of AI in activities such as research, writing, school or professional projects, and data analysis continues to grow, but public trust remains limited. 76% of respondents said they trust AI rarely or only sometimes, while only 21% claim to trust it most of the time or almost all the time.

For companies implementing AI in products and internal processes, this gap between adoption and trust is one of the most important signals in the market. It explains why transparency, traceability of responses, and verification mechanisms become real differentiators, not just compliance elements.

Mercor enters the center of a security breach

Mercor, a company that produces training data for AI, was hit by a security incident. Following the report, Meta paused its collaboration with Mercor, and OpenAI is investigating the incident. The immediate reaction of some major players shows the sensitivity of the data supply chain for AI.

For the industry, the theme is critical. AI infrastructure does not only mean models and accelerators, but also data providers, labeling, and preparation pipelines. Any breach in such a node directly affects trust in the integrity of datasets and in the security of collaboration between platforms and providers.

Perplexity faces accusations regarding conversation sharing

A proposed class action lawsuit accuses Perplexity of having included Meta and Google trackers in its AI search engine and of having transmitted conversations, emails, and other identifiers even when users activated Incognito mode. The Verge cites information that previously appeared in Ars Technica and summarizes the essence of the complaint: users would have been tracked without real privacy protection.

For conversational AI products, this type of litigation is important because it puts direct pressure on how privacy settings are defined and communicated. The difference between the product promise and technical implementation can quickly become legal, reputational, and commercial risk.

Gemma 4 moves to a more permissive license

Google has changed the licensing model for Gemma 4, moving it to Apache 2.0. According to The Verge, previous versions used a criticized proprietary license for being too restrictive, while the new choice is more permissive and much more familiar to developers. Performance improvements for the new version are also emerging.

The change is important for the open model ecosystem because the license directly influences adoption, commercial integration, and the speed at which the community can build around a model. For companies, a standard and predictable license reduces legal friction and increases the chances of experimentation in real products.

China proposes strict rules for "digital humans"

Reuters reports that the Chinese cyberspace authority has published draft rules for the development and use of "digital humans." The proposals require clear labeling of generated content, prohibit the use of other people's personal data without consent, and block interactions with minors that imitate intimate relationships. Additionally, services cannot provide addictive experiences for children, and providers must intervene if signs of self-harm risk appear.

The underlying signal is that avatars, conversational agents, and synthetic interfaces are increasingly moving from the experimental zone into the regulatory zone. For software companies developing products with virtual identity, synthetic voice, or generated human representations, labeling and user protection requirements will become increasingly hard to ignore.

DeepSeek V4 is expected to run on Huawei chips

The next DeepSeek model, V4, will run on Huawei chips. Several major Chinese companies, including Alibaba, ByteDance, and Tencent, are said to have placed consistent orders for the new chips, and DeepSeek has rewritten parts of the code for compatibility with local hardware, in collaboration with Huawei and Cambricon. The model's launch is expected in the coming weeks.

The subject has strategic weight as it shows an acceleration of the Chinese AI ecosystem around a more autonomous stack of hardware and software. If this direction is confirmed, global competition will no longer be fought just between models, but between complete ecosystems capable of operating with less dependence on Western suppliers.

A fake WhatsApp was used in a spyware campaign

Meta has warned of a campaign involving a fake WhatsApp application used as spyware, affecting approximately 200 users, mainly in Italy. The attack did not compromise the official app, infrastructure, or WhatsApp encryption, but relied on social engineering: victims were convinced to install the app outside official stores. Once installed, it could collect messages, contacts, location, and even access the microphone or camera.

For mobile product security, the case is relevant because it shifts attention from technical exploitation to fraudulent distribution and impersonation. Even when the official app remains secure, the ecosystem around the brand can become a major attack surface, especially in contexts where users install apps through links or unverified sources.

A North Korea attributed attack hits a widely used open-source software

Reuters and TechCrunch reported on the compromise of Axios, an open-source library widely used for connecting applications to web services. Attackers introduced malicious code in an update, turning the incident into a supply chain attack with the potential for distribution to millions of environments. Google and security firms have linked the operation to a group associated with North Korea, and analyses have shown that the malware could affect macOS, Windows, and Linux.

This is one of the clearest recent examples of the fragility of invisible software infrastructure. For companies, the message is direct: highly popular open-source dependencies must be treated as critical assets, with integrity checks, monitoring, and rapid response processes, because a single compromised update can escalate into a widespread incident.

Australia and Anthropic sign an agreement for AI safety

Anthropic has signed a memorandum of understanding with the Australian government for cooperation in research related to AI safety and to support the objectives of the National AI Plan. The company also announced partnerships of 3 million AUD with research institutions in Australia for the use of Claude in diagnostics and treatment, education, and research in computing. The agreement includes collaboration with the AI Safety Institute in Australia and the exchange of data and findings about capabilities and emerging risks.

From an institutional perspective, this type of partnership shows how the relationship between states and frontier AI providers is moving towards more structured formulas that combine research, risk assessment, and economic interest. For the industry, it is yet another sign that AI governance is being built in parallel with commercial expansion.

Samsung is discontinuing its messaging app

According to Engadget, Samsung will discontinue the Samsung Messages app in July and replace it with Google Messages. The information indicates a clearer consolidation around the Google messaging platform on Android and a reduction in application duplication at the manufacturer level.

For the Android ecosystem, it is a significant move as it promotes the standardization of messaging experience and features, including around RCS and integration with other Google services. For manufacturers and developers, this means less fragmentation but also a greater concentration of control at the platform level.

OpenAI buys TBPN

OpenAI has acquired TBPN, an online show and media network known especially in the technology and AI space. Reuters and The Verge report that TBPN will maintain editorial independence, and OpenAI sees the acquisition as a way to better articulate its vision and influence the public conversation about the impact of AI. The article also notes that TBPN has a solid audience and a strong presence around interviews with industry leaders.

For the industry, the acquisition shows that the battle for AI is not only fought in labs and products but also in controlling the public narrative. Large companies are beginning to treat media distribution and cultural influence as strategic parts of market infrastructure.

The leak of Claude Code raises new security questions

The source code leak for Claude Code has become one of the most discussed recent breaches in the AI industry. Public summaries show that the problem started with the accidental publication of a source map file in a public npm package, exposing hundreds of thousands of lines of code and nearly 2,000 files. Anthropic stated that the incident was the result of human error and that no customer data or credentials were compromised.

The implication for software companies is clear: in an era of rapid distribution and AI tools for developers, the hygiene of the build and release chain becomes as important as the security of the final application. An incident of this type can expose both intellectual property and sensitive product direction.

Europe attributes a major attack to two hacker groups

CERT-EU stated that a recent breach in the cloud infrastructure of the European Commission was the work of a group called TeamPCP, and the data was subsequently published online by ShinyHunters. Approximately 92 GB of compressed data was extracted from the compromised AWS account, including names, email addresses, and email content. The affected infrastructure served the Europe.eu platform, used by European institutions and agencies, and at least 29 other EU entities could be affected.

The case is relevant due to the combination of compromising an API secret, pivoting in the cloud, and propagating the impact across multiple institutions. For organizations, the incident shows how much risk has shifted around compromised keys, cloud accounts, and software dependencies.

Japan makes physical AI a priority

TechCrunch describes Japan as one of the spaces where "physical AI" is beginning to move from experiment to implementation. The main driver is not enthusiasm for novelty, but the labor shortage and the need for operational continuity in factories, warehouses, and critical infrastructure. The Ministry of Economy, Trade, and Industry in Japan wants to build a domestic physical AI sector and achieve a 30% share of the global market by 2040.

For the industry, the stake is that AI-powered robotics is starting to be seen as productivity infrastructure, not just a technological demonstration. Where demographic pressure is high, advanced automation becomes a direct economic response, and control and orchestration software gains an equally important role as hardware.

Microsoft expands its portfolio of proprietary models

Microsoft AI has launched three new foundational models for text, voice, and image, in a clear step to strengthen its own multimodal stack. MAI-Transcribe-1 covers transcription in 25 languages and is presented as being 2.5 times faster than the Azure Fast offering. MAI-Voice-1 can generate 60 seconds of audio in one second and allows for custom voice creation, while MAI-Image-2 is a model for video generation. The models are available through Microsoft Foundry, and partly through MAI Playground.

For the market, the launch shows that Microsoft is pursuing a dual model: continuing its partnership with OpenAI while also simultaneously developing its own capabilities. This is relevant for enterprise clients as it can offer them more options for cost, integration, and control within the same commercial platform.

Human neurons enter the discussion about the future of computing

Euronews presents Cortical Labs, an Australian startup that claims to have created the first device that allows code to run on living human neurons. The CL1 system combines neurons grown from stem cells with silicon-based hardware, and the company says applications can range from neuroscience and disease modeling to robotics and AI. The use of human cells in computing raises ethical questions, even though some researchers believe that simple networks do not yet pose major issues.

Even though it sounds experimental, the theme is important because it shifts the discussion about efficiency and computing architecture into a completely different area from traditional hardware. So far, the value seems closer to research and very specialized niches, but the direction shows how open the search for new computing paradigms has become.

Nebius builds one of the largest AI centers in Europe

Nebius has announced the construction of an AI facility in Lappeenranta, Finland, with an energy capacity of up to 310 MW, equivalent to about three hyperscale centers. The project will use liquid cooling in a closed circuit to limit water consumption, and the residual heat will be redirected to the local district heating network.

For the European market, the project is important as it shows that AI infrastructure is moving to a scale closer to that of major hyperscalers. Additionally, the combination of high capacity, energy efficiency, and integration with local infrastructure says a lot about what the next generation of dedicated AI data centers will look like.

If you want, I can continue with the second part in exactly the same format for the rest of the links in the list that I haven't included here yet.

Teenagers and their relationships with AI enter an increasingly sensitive area

An article published by Digital Trends shows that teenagers are no longer using chatbots just for homework, quick questions, or specific curiosities. More and more of them are transforming them into constant presences in their digital lives, building routines, recurring jokes, and relationships that include emotional components, roleplay, and forms of attachment that exceed the classic idea of software tools. The material cites data from Common Sense Media indicating that 72% of teenagers have used AI companions, and 33% say they have used them for friendship or companionship.

The article emphasizes that the phenomenon can no longer be treated as a simple extension of chatbot culture. In practice, the interaction becomes more personal and more constant, and for some users, AI begins to occupy a social role. This significantly changes how these products need to be understood: not just as conversational interfaces, but as experiences capable of creating emotional dependence, attachment, and repetitive behaviors.

For the software industry, the implication is important. When an AI product moves from the utility zone to the relationship zone, design, safety, and moderation standards need to be rethought. Especially in the context of minor users, the question is no longer just how good the model is, but how responsible the product delivering it is.

LinkedIn and the controversy over scanning browser extensions

The Next Web reports a controversy regarding how LinkedIn allegedly scanned installed browser extensions, in a list that reportedly reached 6,167 extensions in February 2026. The article states that, according to cited tests, the checks were still active at the beginning of April and that the scope of monitoring has increased significantly in recent years, from 38 extensions in 2017 to 461 in 2024, then to over 6,000 in 2026.

The stakes are not just about the extensions themselves, but about the type of inferences that can be drawn from them. The material notes that the list may have included tools associated with job searching, religious practices, neurodivergent conditions, or political interests, i.e., areas considered sensitive within GDPR. Even though LinkedIn rejects the accusations and states that it monitors extensions that scrape data without consent and that it does not use the information to deduce sensitive data about members, the discussion remains open.

For companies operating large platforms, the case is relevant because it shows how quickly a measure presented as technical protection can become a privacy and trust issue. For users and product teams, the difference between security, platform control, and excessive signal collection becomes increasingly difficult to manage without clear transparency.

Chrome turns security updates into a more visible signal

A Yahoo Tech article draws attention to a new message that may appear in Google Chrome indicating the existence of a pending security update. The central idea is simple: if the user sees that message, the browser is no longer running the most secure version available and remains exposed until the update is installed. The article insists that the alert does not appear arbitrarily, but is linked to a real update that needs to be applied.

The subject comes in a context where Google has issued several recent warnings regarding security risks for Chrome, including around actively exploited vulnerabilities. For this reason, the message should not be treated as a simple maintenance notification, but as an operational signal that the exposure window is still open.

For individual users, the recommendation is trivial but important: the update should not be delayed. For companies, the implication is broader, as the browser is one of the most exposed components in the daily work stack. When critical updates are not installed quickly, the attack surface increases directly at the end-user level.

Microsoft links its AI strategy more clearly to business value

Microsoft describes a shift in focus in its AI strategy. After the internal reorganization in March 2026, Mustafa Suleyman has shifted his attention almost exclusively to the development of frontier AI and what he calls "superintelligence," but not in an abstract or philosophical sense, but in one oriented towards concrete value for enterprise clients, developers, and users.

In the same context, Microsoft launched MAI-Transcribe-1, a transcription model for 25 languages, built for challenging scenarios such as background noise, poor-quality audio, and overlapping speech. The company claims that the model is more cost-efficient in terms of GPU and positions it alongside MAI-Voice-1 and MAI-Image-2 in AI Foundry and Microsoft AI Playground, marking a broader commercial availability for these models.

For the software market, the signal is clear. Microsoft is trying to turn its AI into an operational advantage, not just a branding bet. This matters for companies choosing enterprise platforms, as it means more options for integration and a more direct competition among model providers on cost, speed, and practical utility.

Apple at 50: a big anniversary, and even more pressure on AI

On its 50th anniversary, Apple is presented in several public summaries as a company entering an important anniversary stage, but also at a moment of strategic pressure. The focus is no longer just on the legacy of Macintosh, iPhone, or Apple Watch, but on the question of whether Apple can deliver a new product cycle in AI that is compelling enough for investors and the market.

Apple's future is read through the lens of delays surrounding Siri and the accelerated competition in the AI space. In other words, the anniversary is not just a celebration, but also a point of evaluation: how quickly can the company transform AI into a coherent product experience, at scale, and with an impact similar to its major historical launches.

For the industry, the theme is relevant because Apple continues to influence how technology reaches the general public. If Apple seriously accelerates in AI, the effects will be seen quickly in personal assistants, on-device processing, integration between hardware and software, and implicitly, in user expectations regarding everyday digital products.

Anthropic and the security incident surrounding Claude Code

A recent incident exposed source code from Claude Code, after a source map file was accidentally published in a public npm package. Available information shows that the exposure included hundreds of thousands of lines of code and a large number of internal files, raising concerns in the developer community.

Anthropic stated that the incident was the result of human error and that no user data or sensitive authentication information was compromised. However, the exposure of the code raises questions related to the protection of intellectual property and internal software publishing processes.

For tech companies, the case is a clear example of the importance of security in the development pipeline. In an ecosystem where package distribution is rapid and automated, a minor error can have a major impact on code and on trust in the product.

Europe is massively investing in AI infrastructure

A Dutch company has announced the construction of one of the largest AI data centers in Europe, with a capacity of hundreds of megawatts. The project includes advanced cooling technologies and solutions for reusing energy generated by servers.

The investment reflects the growing demand for infrastructure capable of supporting the training and running of AI models at scale. Data centers are becoming critical elements in the global competition for AI.

For companies, the development of these facilities in Europe can reduce dependence on infrastructure from other regions and can accelerate AI adoption in local projects, providing faster and more efficient access to computing resources.

Synthesis made with the help of a monitoring flow provided by Control F5 Software.