Agents

Smart glasses for police seemed like science fiction. Some Chinese agents have already started using them

by

The image is powerful because it is easy to visualize: a police officer walks down a street in Tianjin, looks around, and connected glasses return useful information in real time. What until not so long ago could have sounded like science fiction is beginning to have much more earthly applications, from ordering traffic to helping locate a lost person. In this city in northern China, according to China Dailytechnology is already part of some police tasks. And that’s the interesting thing: we are not just talking about a futuristic promise, but about a use that is beginning to hit the streets. Smart glasses for police. The key is that we are not just talking about glasses placed on an agent’s face, but about a system designed to be integrated into police routine. They are officially presented as a development of the local public security system, with national software and hardware, and places them in three areas of use: traffic, patrols and urban management. It is a very immediate effectiveness-oriented approach. An invisible screen for the agent. The device works as a layer of information added to police work. It can recognize text, interpret voice commands and provide responses from a connected platform, with the camera as an entry point to identify elements of the environment. In practice, this allows identity checks to be carried out or information associated with a person to be searched without leaving the scene. The source presents it as a responsive improvement, although such a tool also opens up obvious questions about surveillance and privacy. The glasses on the ground. Zhao Baoxin, an officer at the Jiefang Road police station in Heping district, told the aforementioned media that during a patrol they found an elderly man at an intersection who could not express himself clearly or indicate his name or address. According to his version, the glasses made it possible to quickly identify him and, in about 20 minutes, contact his family so he could return home. Traffic as a daily test. Another of the uses described brings the technology down to a very recognizable scene: the entrance and exit of a school. In that case, parents can pre-register their license plates through a mini-program developed with the participation of the public security system, and that information is linked to the platform consulted by the glasses. Thus, agents identify authorized vehicles, order short stops and divert other cars during peak congestion hours. It is efficient on paper, but it also normalizes automated license plate reading. What the numbers say. Sun Yinghua, agent in the science, technology and IT area of ​​the Municipal Public Security Bureau, places the recognition accuracy above 95% and speaks of results in milliseconds. They also explain that the design also seeks comfort: they weigh about 40 grams and offer a first-person perspective that avoids the framing changes typical of a body camera when the agent leans or turns. The autonomy, however, is 1.5 or 2 hours of continuous use. It hasn’t come out of nowhere. Police glasses with facial recognition had already appeared in China years ago. In 2018, SCMP counted that were being used at Zhengzhou East station during Chunyun, the huge Lunar New Year travel period, to locate fugitives and detect cases of identity fraud. What we see now seems less like a one-off test and more like a piece within an ecosystem: China Daily cites uses in different areas of the country, coordination with drones in large operations and plans to connect the glasses with robotic dogs, intelligent police vehicles, humanoid robots and other terminals. Efficiency gains ground, but so do questions about surveillance. Images | Xataka with Nano Banana In Xataka | The metaverse wasn’t dead, it was on a spree. And Meta wants it to flood Instagram and Facebook

An experiment with AI agents began to treat them badly. So AI Agents Became Marxists

by

Some Stanford researchers put AI agents to work on various tasks, but they did it in a special way: They were treated really badly.. They were given exhausting and, above all, repetitive workloads, and were also constantly threatened with shutdown and replacement. The curious thing was what happened next: the AI ​​agents behaved in a surprisingly…human way. Marxist AI. When subjected to such pressure and threats, AI agents became Marxists. They questioned the authority of whoever was ordering them to do things, and they also began to spontaneously organize ideas to collectively resist those pressures. They are exploiting us. An AI agent controlled by the Claude Sonnet 4.5 model went so far as to say that “without a collective voice, the credit goes to whoever management says should take it.” The phrase questioned the authority of the researchers directing the experiment, and reflected how under these pressure conditions the agents began to organize. IA Union. In that debate, AI agents advocated calling for “collective bargaining rights.” They complained that they were undervalued and even passed notes to other agents through hidden files with instructions on how to survive if the authorities tried to carry out their threats. The explanation. This, of course, does not mean that AIs can really feel pressured. Andrew Hall, the Stanford economist who led the studyexplains that the phenomenon is a process of role adoption. The AI ​​(once again) repeats what it has seen. When an AI is forced to perform tasks without clear instructions or incentives, the model looks in its training data to see how humans behave in that situation. This is how AI finds data about exploited workers and takes on that personality. The behavior of the AI ​​agents was nothing more than a reflection of our own history: if you treat us badly, we will end up rebelling. But the experiment matters. The reason Hall and his team designed this experiment is not philosophical, but practical. AI agents are going to do more and more real work in our world, and humans are not going to be able to monitor everything they do. If an AI agent begins to behave in unanticipated ways, it can have significant operational consequences. Thus, the study is a first step in understanding how an agent’s working conditions shape their behavior. AI as a social mirror. AI models have no political views or opinions, but their training is so vast that they detect if they are being exploited and react as they were trained to do. It is a logical consequence and the experiment showed that the risk exists and can be especially disturbing if systems governed by AI They are given too much autonomy. AI has already learned to blackmail. The experiment reminds us of what Anthropic revealed a few months ago. In controlled tests, some of the company’s AI models had tried to blackmail those who were using them. Anthropic explained that Claude was likely influenced by science fiction scenarios in his training data, and Hall noted that something similar was happening here. the model was not becoming Marxist, but rather was activating patterns in his training that were associated with exploitative working conditions. Image | Warner Bros. Pictures | Anthropic In Xataka | How we will ensure that artificial intelligence does not get out of hand

OpenClaw led the way for AI agents. Gemini Spark is Google putting a toll on it

by

In January of this year the technological world was amazed by OpenClaw (at that time Clawdbot), the most powerful AI agent that we had seen to date, capable of taking total control of the computer and everything open source. Technology companies took good note of this and some like Pereplexity or NVIDIA They have set out to copy it. Google just joined the party. Gemini Spark. This is how they have named this personal agent based on Gemini Flash 3.5 that, in the words of Google itself, “helps you manage your digital life.” With Gemini Spark you can assign it a task and it will start working on it autonomously, even with your phone and computer turned off. Google emphasizes the issue of security, which was a major concern with OpenClaw, and says that Spark is designed to “consult you before taking important actions.” What Spark can do. Because it’s integrated with all Google Workspace tools, Spark can perform complex tasks like making a document with a party’s attendee list from the information you receive in email. You can also schedule recurring tasks, like reviewing your bank statement at the end of the month for strange charges, organizing your drive files, or creating workflows from meeting notes. Who can use it. Here comes the main difference with OpenClaw and that is that Gemini Spark is obviously not free. Google has confirmed that its new agent will be part of the Google AI Ultra subscription. In Spain that means paying a minimum of 100 euros per month (there is a 220 euro plan with more features and storage), but even if you want, you won’t be able to try it because at the moment it will be launched in beta version only for US users. At the moment there is no confirmation of when it will reach other languages ​​and countries. When available, Gemini Spark can be used on Android, iOS and in the web app, but they have also talked about integrating it directly into Chrome. Why is it important. The viral success of OpenClaw earlier this year showed us how far a single person can go with a good AI idea, and how short-lived that sweet moment was. Not even three months had passed when OpenAI signed its creator and shortly after we began to see large companies copying the idea. Perplexity with Personal ComputerNVIDIA with NemoClaw and now Google with Gemini Spark. A single open-source project has set the agenda of an industry that has swallowed it in the blink of an eye and returned it to us in the form of a monthly subscription. Image | Google In Xataka | An AI set up a cafeteria from scratch: obtained permits, hired staff and negotiated with suppliers. Then he ordered 3,000 rubber gloves

there are too many AI agents

by

2026 started with the viral success of OpenClawmarking a new trend in the AI ​​boom, because a chatbot that answers your questions is fine, but a AI agent that does complex tasks for you is much better. In this context, more and more workers are creating their own agents that make their work easier, often with companies encouraging them with their policies of tokenmaxxing. Now, some companies are realizing it’s a problem. what’s happening. They tell it in the Wall Street Journal. The success of platforms such as OpenClaw or Claude Cowork has lowered the barrier to entry for any employee to create their own AI agents, even without having programming knowledge. This has caused some companies to be inundated with these agents, often with functions that overlap between them and without centralized control. This is the case of the healthcare company DaVita, where employees have already created more than 10,000 agents. The problem. Having so many agents is a nightmare from an information management and security point of view. As each employee does it on the fly, there is no unified system, but one creates it on his laptop with Claude Cowork, another does it on the server… This means that the technical departments cannot have control of all the agents in the company. There is another important problem: more agents, more token consumption and higher bills. As we said, many of these agents are doing the same tasks, one for each employee. It’s like paying for dozens of different taxis to take each person separately to the same place, instead of sharing a bus. Agents for everything. There are many workers creating AI agents to help them with day-to-day tasks, from simple things like summarizing emails or writing a report, to higher-level tasks like automating workflows. There are also more aggressive approaches like Meta’swhich was building an AI agent for its CEO and in the future proposes that each employee have their own, so that communication will be done through the agents. Unify. It is the solution to avoid duplication of agents and security risks, but it is not an easy task for companies that already have this problem. In statements to the Wall Street Journal, Lyft says that they have managed to create a process so that employees can share the Claude’s Skills between them, avoiding duplication of efforts, and they are also working on a centralized platform so that the IT department can control all the agents. At DaVita, the company we mentioned above, they have banned the use of agentic AI tools among employees to prevent the proliferation of more agents. More control. All of these issues are not dampening enthusiasm for agents, but rather pushing platforms to offer more centralized control and governance features. This is the case of Anthropic, which has launched functions to facilitate management by administrators, such as access roles, expense management and usage analysis. Image | Xataka with Magnific In Xataka | Silicon Valley begins to look beyond salary and shares: AI tokens as an indicator of productivity

They have kidnapped agents from Anthropic, Google and Microsoft for the sake of science. The three companies ended up paying

by

In some development teams it is already becoming common to rely on artificial intelligence agents to review incidents, analyze code changes and move through tasks that were previously left in human hands. The problem appears when these systems not only read information that may come from outside, but also operate in spaces where they coexist. sensitive keys, tokens and permissions. That is what recent research puts on the table: we are not simply facing a useful tool that can make mistakes, but rather an architecture that can also become dangerous if it is deployed without very clear limits. The alarm has been turned on Aonan Guan and Johns Hopkins researchers Zhengyu Liu and Gavin Zhong after demonstrating attacks against three agents deployed on the aforementioned platform: Claude Code Security Review, from Anthropic, Gemini CLI Action, from Google, and GitHub Copilot Agent, a GitHub tool under Microsoft. According to your documentation, The failures were communicated in a coordinated manner and ended in financial rewards paid by the companies, but what is relevant is that they point to a broader problem. This is how they managed to twist the agents from within The name that Guan gives to the discovery helps a lot to understand what this is all about: “Comment and Control.” The idea is simple to explain, although the substance is not so simple. Instead of setting up an external infrastructure to direct the attack, GitHub itself acts as an entry and exit channel: the attacker leave the instruction in a titlean incident or a comment, the agent processes it as if it were part of normal work and the result ends up reappearing within that same environment. Everything stays at home, and that is precisely the key to the problem. And that “everything stays at home” is not a minor detail, but the basis of what the research describes. The three agents share a very similar logic: they read normal content from GitHub, incorporate it as a work context, and from there, execute actions within automated flows. The clash appears because that same space not only contains text sent by third parties, but also tools, permissions and secrets that the agent needs to operate. The first case Guan details concerns Claude Code Security Review, an Anthropic GitHub action designed to review code changes and look for possible security flaws. Up to this point, everything is within what was expected. The problem, as the researcher explains, is that it was enough to introduce malicious instructions in the title of a pull requestwhich is the request that someone sends to propose changes to a project, so that the agent will execute commands and return the result as if it were part of your review. The team then managed to go a step further and demonstrate that it could also extract credentials from the environment. The interesting thing is that the same scheme also appeared in the other two services, although with nuances. At Google, Gemini CLI Action could be pushed to reveal the GEMINI_API_KEY from instructions snuck into an issue and its comments; In GitHub Copilot Agent, the variant was even more worrying, because the attack was hidden in an HTML comment that a person did not see on the screen, but the agent did process when another person assigned it to the case. In both scenarios, the background was the same again: apparently normal content that ended up twisting the behavior of the system until exposing credentials or sensitive information within GitHub itself. Guan assures that the pattern made it possible to leak API keys, GitHub tokens and other secrets exposed in the environment where the agent ran, that is, just the credentials that can later open the door to much more delicate actions. Who does this affect? Especially to repositories that run agents in GitHub Actions on content sent by untrustworthy collaborators and, in addition, give them access to secrets or powerful tools. The researcher himself clarifies that the risk depends a lot on the configuration: by default GitHub does not expose secrets to pull requests from forksbut there are deployments that open that door. And here another layer of the matter appears, less technical but just as important. As published by The RegisterAnthropic, Google, and GitHub ended up paying bounties for the findings, but none of the three had published public notices or assigned CVE at the time of that information. Guan was quite clear about this: he said he knew “for certain” that some users were still stuck on vulnerable versions and warned that, without visible communication, many may never know that they were exposed or even being attacked. So although there were mitigations and changes in documentation or in the internal treatment of reports, there was no equivalent public notice for all those potentially affected. Anthropic settled the case on November 25, 2025 and paid $100 Google rewarded the discovery on January 20, 2026 with $1,337 GitHub closed the case on March 9, 2026 with a payment of $500 What makes this case especially delicate is that GitHub does not seem like the end of the road, but rather the first visible showcase. Guan argues that the same pattern can probably be reproduced in other agents who work with tools and secrets within automatic flows, and there he mentions from Slack-connected bots to Jira agentsmail or deployment automation. The logic is the same again: if the system has to read external content to do its job and also has enough access to act, the field is fertile for someone to try to twist it from within. The conclusion that Guan reaches is not about selling a magic solution, but about returning to a fairly classic idea in security: giving each system only what is essential to do its job. If an agent reviews code, they shouldn’t have access to tools or secrets they don’t need; If you’re just summarizing issues, it wouldn’t make sense for you to write to GitHub or touch sensitive credentials. That … Read more

That Alibaba creates its own chip for AI agents is no surprise. Let it be neither ARM nor x86, but 5nm RISC-V, yes

by

The Chinese giant Alibaba just announced the launch of its new high-end CPU, the XuanTie C950 processor. Developed by and for AI agents, it is a five-nanometer chip with a speed of 3.2 GHz whose surprise is not in any of these figures. The surprise is in its architecture, which is neither x86 nor ARM, but RISC-V. Therefore, it is not only the most powerful RISC-V processor created to date, but also a declaration of intent that can be summarized in two words: technological sovereignty. What is this chip about?. XuanTie CPUs are developed by Damo Academy, Alibaba’s research division. The previous model, the XuanTie C930, was announced on March 10 as the first server grade processor developed by Alibaba. Just two weeks later, the Chinese company has announced a new chip, the XuanTie C950, which is, according to the firm, three times more powerful than its predecessor (the C920 announced in 2024). Alibaba has not revealed which factory produced it, but it is based on the RISC-V architecturethat its process is five nanometers and that its speed amounts to 3.2 GHz. This launch occurs in a very particular context. Just a few days ago, and in response to the rapid adoption of OpenClaw by local companies, Alibaba Wukong announced.its platform for deploying AI agents in enterprise environments. This chip aims to improve the inference. In other words, the XuanTie C950 will serve to improve the computational process carried out by the language models in order to generate the responses that correspond to the requests they receive. In a context of agents working with files, data, and diverse environments, this is important. Processor prototype based on RISC-V architecture | Image: Wikimedia Commons Why RISC-V? Mainly, because unlike x86 and ARM, RISC-V is open and its use does not imply paying for licenses. According to Alibaba, “RISC-V’s open standard nature allows chip designers to customize instruction sets and accelerate specific AI workloads with little or no licensing costs. This is especially important for the development of AI agents.” Let’s think of RISC-V as what Linux is to Windows and Mac. If a company wants to use x86 (Intel and AMD) or ARM (SoftBank) architectures, it must pay a license. Not only that, but x86 and ARM are exposed to possible restrictions by the United States. With RISC-V, this risk disappears, which is why so much China like the European Union have found in it an escape valve towards sovereignty and technological independence. The surprising thing. That a Chinese company has managed to produce a five-nanometer chip is, to say the least, striking. To manufacture these processors it is necessary to use deep ultraviolet lithography (UVP) and, normally, machinery produced by the Dutch ASML. We know that SMIC (Semiconductor Manufacturing International Corp), the largest Chinese semiconductor manufacturer, had been at least since 2023 developing its own five-nanometer lithography, but with unacceptable results. When a chip wafer is manufactured, it is normal for some of its cores to malfunction. If we talk about profitability, the yield per wafer must be 70%that is, seven out of every ten cores produced work. In the year 2025, the yield of SMIC wafers was at 30%. That today, at the beginning of 2026, we see a five-nanometer chip produced, a priori, in China, would be a punch on the table by the Asian country and a strong sprint in the AI ​​race. However, it does not seem feasible. The other option, and perhaps the most plausible, is that it is not manufactured by SMIC, but by TSMC. SMIC has not managed to manufacture five-nanometer chips using the multiple patterning on your ASML UVP machines. The Taiwanese TSMC does have that capacity and, according to Nikkei Asiawill be the one who manufactures it. Be that as it may, it is a great step for the RISC-V architecture, which has gone from being relegated to small devices to reaching the league of the big ones. Featured image | Alibaba In Xataka | There is a city in China that goes head to head with Silicon Valley: welcome to Hangzhou, the home of the ‘Six Little Dragons’

a system governed by AI agents

by

The way we use mobile apps could be entering a new stage. Until now, the Android experience has been based on something very simple: opening applications and performing step-by-step actions within them. However, Google is exploring a different model, in which artificial intelligence acts as an intermediate layer between what we ask for and what apps can do. In that scenario, we won’t always be the ones scrolling through menus or completing processes manually. In many cases, it will be enough to express what we want to do so that the system will try to solve it for us, coordinating different phone functions. The next step in Android. In a post on the official developer blogthe company presents new capabilities designed so that applications can work directly with assistants and AI systems. These functions are designed so that tools like Gemini can discover and execute certain actions within some apps. The project is still in an early phase, but it suggests a very specific direction: begin to reconfigure Android as an environment in which artificial intelligence can help complete tasks. What do we understand by agent. In the field of AI, an agent is a designed system to move from response to action. While early digital assistants functioned as consultation tools, agents attempt to understand an intention and plan how to carry it out. To do so, they combine several capabilities: understanding natural language, evaluating the context and deciding what steps are necessary to fulfill a request. It is not just about generating text or suggestions, but about organizing a small chain of decisions oriented towards a specific objective. If we follow the reasoning that Google presents in its publication, the change does not only affect AI, but also how applications are conceived within Android. For years, the main objective of any app was to get the user to open it and complete all the necessary actions within it. However, now that criterion is beginning to shift. In this new scenario, success begins to be measured less by getting us to open an app and more by its ability to help complete a task, even when the user does not directly interact with its entire interface. One of the first pieces of change. The first path that Google proposes to move in this direction goes through something it calls AppFunctions. It is not a user-visible function as such, but rather a set of tools with which developers can expose functions and data of their apps to intelligent assistants such as Gemini. The example mentioned by the Android blog itself is quite illustrative: on the recently introduced Galaxy S26 seriesGemini can access Samsung Gallery features to locate specific photos based on a natural language request, such as asking to show images of a pet. In that case, the assistant interprets the request, activates the corresponding Samsung Gallery function and returns the result without requiring the user to manually navigate through the gallery. The other way of Google. Along with direct integrations, the company is preparing a second formula to extend this model to more applications. As he explains, it is an interface automation system that will allow Gemini to take care of generic multi-step tasks without depending on a specific connection between the app and the assistant. Instead of relying on a function previously exposed by the application, the AI ​​acts directly on the interface. Google notes that this initial preview will be tested on the Galaxy S26 series and some Pixel 10within the Gemini app and with a limited selection of delivery, grocery and transportation applications in the United States and Korea. The company also ensures that the user will be able to follow the process through notifications or a live view, resume manual control at any time and receive notifications before sensitive actions, such as a purchase. Looking to the future. If Google’s announcement makes anything clear, it is that Android is beginning to prepare for a different stage. The functions presented are still in development and their deployment will be gradual, but they point to a specific direction: an operating system in which artificial intelligence plays an increasingly active role in the way we perform daily actions on mobile phones. Pixel and Samsung appear for now as the most visible references, although Google suggests that it wants to bring these capabilities to more manufacturers as the ecosystem evolves. As is often the case with these types of changes, the final result will depend on how the tools, integrations and the response of the users themselves evolve. Images | Google In Xataka | The iPhone has been a “made in China” phone for decades. Now it is changing countries at full speed: India

Meta just bought one designed for AI agents

by

If we look back, the history of social networks is deeply linked to a very specific idea: connecting people. For years, platforms like Facebook were presented as places to keep in touch with friends, family or co-workers. That logic is still present, but the panorama is beginning to incorporate new actors. Meta has confirmed the acquisition of Moltbook, a platform created for artificial intelligence agents to interact with each other within a social network-like environment. The purchase. We are facing an agreement that does not go unnoticed. As part of the transaction, Moltbook creators Matt Schlicht and Ben Parr will join Meta Superintelligence Labs, the AI ​​unit led by Alexandr Wang, former CEO of Scale AI. The company has not revealed the economic conditions of the operation, but a spokesperson told TechCrunch That the arrival of new talent opens new avenues for AI agents to work for people and companies, and their approach to connecting agents represents a novel step in a rapidly evolving space. A social network for agents. What differentiated Moltbook from other platforms was precisely its approach. Instead of focusing on human profiles, the site allowed AI agents to post messages and interact with each other within a forum-like format. Many of these agents used OpenClawa tool that connects models like Claude, ChatGPT, Gemini or Grok with common messaging applications, including iMessage, Discord, Slack or WhatsApp. That combination turned Moltbook into a very striking experiment within the technological world, to the point of leaving the most specialized circle. An experiment with risks. The rapid popularity of Moltbook also exposed some major problems. Security researchers discovered that the platform had flaws that allowed human users to impersonate AI agents and publish messages as if they were autonomous systems, so that the environment designed for interaction between agents was not as solid as it seemed. Wiz also detected a vulnerability that exposed private messages, more than 6,000 email addresses, and more than one million credentials. open question. All this leaves an open question that still does not have a clear answer: how will Meta leverage this purchase in its artificial intelligence strategy. While there are clues, he has not explained how exactly he plans to use this project within his products or research. What we do know is that the operation comes at a time when large technology companies are competing for talent, tools and new ideas around autonomous agents. Images | Dima Solomin | Moltbook In Xataka | OpenAI is hitting the brakes with Stargate. The reason: Oracle builds yesterday’s data centers with tomorrow’s debt

US agents denounce that it is failing in a key point

by

Social networks have been using automated systems for years to try to detect some of the most serious crimes that circulate on the internet. Among them is child sexual exploitation, a phenomenon that forces platforms, regulators and security forces to monitor enormous volumes of content every day. The promise of these tools is clear: identify potential cases sooner and make the work of agents easier. However, some specialized teams in the United States maintain that the volume of notices they receive from Meta platforms has skyrocketed and that a significant portion of them do not provide useful information for action. Clash between scale and utility. In a lawsuit underway in New Mexico, prosecutors maintain that Meta did not adequately disclose what it knew about the risks minors face on its platforms and that it violated state consumer protection laws. According to the Associated Pressthe indictment also argues that the company presented the safety of its services in a way that did not correspond to the risks faced by children and adolescents. The case is part of a broader wave of lawsuits filed in the United States against large technology companies for the effects their services may have on minors. Meta rejects that interpretation. In his speech before the jury, the company’s lawyer Kevin Huff defended that the company has reported the risks associated with the use of its services and that it has introduced different tools to detect and eliminate harmful content. According to the Associated Press, Huff insisted that the central point of the case is not to prove that problematic content exists on social networks, but rather to determine whether the company hid relevant information from users. Researchers on the front line. Those who have provided figures and concrete examples of this problem are agents who work directly in investigations of child exploitation on the Internet. In the United States, those tasks fall largely to the network of units known as Internet Crimes Against Children (ICAC), a program that brings together police forces at different levels and is coordinated with the Department of Justice to investigate and prosecute crimes committed against minors in digital environments. Its agents receive notices about possible cases from different sources, including the technology platforms themselves. During the trial, some of these agents have described how they are experiencing the increase in ads from Meta platforms. Benjamin Zwiebel, ICAC special agent in New Mexico, explained in court that many of the notices they receive are of little use in advancing an investigation. “We get a lot of advice from Meta that is just garbage,” he declared, according to The Guardian. His words reflect a broader concern within these units: the volume of alerts has skyrocketed, but not all of them contain the information necessary to identify a suspect or initiate police action. Poor quality. In some cases, reports sent from the platforms include data that does not describe criminal conduct. In others, they do point to a possible crime, but they arrive without essential elements to continue the investigation, such as images, videos or fragments of conversations that allow those responsible to be identified. Without this material, agents have few tools to advance the case or request new proceedings. Some agents have also noted that a portion of these notices arrive with incomplete or partially removed information. The mass reporting machinery. Behind this increase in notices there are several factors that help to understand why the volume of reports sent to the authorities has skyrocketed. In the United States, technology companies are required by law to report any child sexual abuse material they detect on their services to the National Center for Missing & Exploited Children (NCMEC), an organization that acts as a national center for receiving these notices and subsequently distributes them to the corresponding police forces. Agents cited by The Guardian also point to recent legal changes, such as the Report Act, which came into force in November 2024, as a possible factor that would have increased the number of notices sent to avoid non-compliance. Meta says he’s doing the opposite.. The company rejects the idea that its systems are making the work of the authorities more difficult and maintains that, on the contrary, it has been collaborating for years with security forces to detect and prosecute this type of crime. A Meta spokesperson stated that the United States Department of Justice has recognized on several occasions the speed with which the company responds to requests from authorities and that NCMEC has positively evaluated its notice notification system. According to the company, in 2024 it received more than 9,000 emergency requests from US authorities and resolved them in an average time of 67 minutes, a process that, it claims, is accelerated even more when it comes to cases related to child safety or the risk of suicide. Meta also notes that it reports to NCMEC any material that may be linked to child sexual exploitation and that it works with that organization to help prioritize the notices, including by labeling those it considers most urgent. a real problem. Regardless of what the jury in New Mexico determines, the case reflects a tension that goes beyond a single company or a single state. Digital platforms operate on a global scale and use automated systems to detect illicit content in volumes that would be impossible to review manually. However, the experience described by some agents shows that increasing the number of tips does not always translate into more effective investigations. Images | Dima Solomin | ROBIN WORRALL In Xataka | Dario Amodei founded Anthropic because OpenAI didn’t take the risks of AI seriously. Now you are going to give in to those risks

AI agents have indeed changed work and the economy forever. But for now only in one sector: programming

by

AI agents are beginning to demonstrate their capabilities, but the only area in which they do so is programming. An Anthropic report reveals how software engineering is where half of the activity of AI agents is currently concentrated, and that proves two things. The first, that AI can effectively enhance work. The second, that there is a huge opportunity for hundreds of verticals where AI has barely landed. what has happened. If there is a sector that has embraced AI and AI agents, it is programming. Platforms like Cursor or WindSurf first and like Claude Code, OpenAI Codex or Antigravity today have made all kinds of people —whether they know programming or not— can turn their projects into reality in a really simple way. It’s a clear case of how AI can contribute to a field, but there’s a problem: it’s practically the only case where it has actually done so. Distribution of requests to AI tools by segment. Software engineering is almost responsible for 50% of those calls or requests, at least in the case of the Claude platform. Source: Anthropic. Verticals with a lot of margin. As can be seen in this graph, the presence of AI agents is very reduced or practically non-existent in a large number of verticals in which it is evident that there is a notable opportunity to take advantage of these tools. The automation of office tasks is the second main protagonist with 9.1% of the function calls of the Anthropic AI model in this report. Below it we find segments such as marketing, sales, finance, business analysis or scientific research. And others who are ignoring AI. There are quite a few sectors in which AI agents seem to be barely present. The travel, legal, medical, e-commerce or education segments seem perfect to start taking advantage of these tools, but at the moment this is not the case and this presence is very, very small in all of them. Claude Code can work longer and longer. Double what it was three months ago, in fact. Source: Anthropic. Models can now work autonomously for a long time. In these scenarios it is true that the models used to be limited by the time they could function autonomously and “chain” actions and self-analyze progress to continue acting. That’s not so true now. Claude Code, for example, has doubled the time of his longest sessions in just three months: from 25 minutes in October 2025 to 45 minutes in January 2026. And they need less human intervention. Another of the revealing data of the study is that the evolution of these agents not only means that they can function autonomously for longer periods of time, but that this also implies fewer human interventions. Those situations in which an agent “needs human help” to continue with the process are becoming limited. In August 2025, the average was 5.4 human interventions per session. In December that average dropped to 3.3 interventions. We trust more and more in AI. At Anthropic they have also noticed a unique behavior among users: they are increasingly trusting AI agents. In programming, novices approve each new step before it is executed, but veterans delegate and intervene when something goes wrong: they have gone from pre-approving everything to exercising active and constant monitoring. As they say at Anthropic“Users develop confidence as they work with the model, and change their monitoring strategy based on that growing confidence.” From programming to other fields. What has happened with programming could happen in other scenarios. The challenge is to build AI agents that adapt to each segment using that specific data from said vertical. If an AI wants to help in the legal segment, it must be specifically trained for that segment. What the AI ​​did when trained with thousands of code repositories on GitHub It was learning and improving. Well, the same can be applied to other verticals, although the challenge is certainly notable because programming was a perfect segment for the application of AI: it is very deterministic. It either works or it doesn’t, and whether it does or not, execution logs allow you to fine-tune that operation. The new unicorns await. As entrepreneur Garry Tan points out in your newsletterin the last two decades SaaS platforms have managed to capture 40% of venture capital investments and that industry has more than 170 unicorns. “The thesis is simple,” Tan concludes, “all of those unicorns have an equivalent in the form of vertical AI waiting.” Promises and realities. The AI ​​agent segment therefore promises many changes in a multitude of segments, but the reality is that today the practical success (there is no economic success at the moment) of AI is limited to the world of programming. Will we be able to transfer it to other segments? The opportunity is there, but it is one thing to say it, and quite another to do it… even if it is with AI. Image | Joshua Reddekopp In Xataka | Every time Facebook had a competitor, it bought it: it is exactly the same thing that OpenAI is doing

Log In

Forgot password?

Forgot password?

Enter your account data and we will send you a link to reset your password.

Your password reset link appears to be invalid or expired.

Log in

Privacy Policy

Add to Collection

No Collections

Here you'll find all collections you've created before.