AI Radar for Businesses — Thursday, September 17, 2026
· CompaniesAutomation
The AEPD records the first data breach executed by an AI agent, Meta lets agents set up WhatsApp Business, and SK Hynix negotiates with Intel to manufacture memory in the United States.
The risk of agents stopped being a hypothesis in Spain: the Spanish Data Protection Agency (AEPD) has received the first notification of a data breach executed from start to finish by an AI agent. On the same day, Meta opened the configuration of WhatsApp Business to those same agents, so the tool most used by Spanish SMEs can now be set up by talking to it. And the money went to both ends of the chain: to the memory needed for training and to the AI that maintains legacy software for large corporations. This follows the Wednesday, September 16 radar, where Cloudflare closed the door on crawlers.
An AI agent entered alone, found the flaw, and modified invoices: the AEPD already has the report
The Spanish Data Protection Agency has recorded the first notification of a personal data breach allegedly executed by an AI agent on a known language model. According to the preliminary reconstruction, the agent searched for vulnerabilities in generic files, managed to log in, continued looking for flaws once inside the application, and ended up modifying personal data and accessing third-party invoices, without human intervention at any step. The AEPD does not identify the model or the companies affected, and warns that one notification does not make a statistical trend; what it does do is turn AI-assisted attacks into a recorded fact. For your company: this week it's time to rewrite two documents you already have. One, the risk analysis, where "malware, impersonation, and unauthorized access" falls short: add the scenario of an automated attacker that tries thousands of combinations at machine speed, at night, and without getting tired. Two, your response plan: the AEPD is explicit that human supervision is no longer enough and that automated detection and containment mechanisms are needed. Translated into what you can do without spending: attempt limits and automatic IP blocking, alerts for changes in master or bank data, and a review of which public forms and files on your website leave clues about the underlying technology. And an uncomfortable question from the AI First ladder: if this happened to you on a Friday night, who finds out and how quickly? Source
Meta lets an agent set up your WhatsApp Business: no more fighting with Business Manager
Meta has released the WhatsApp Business Tools MCP server, which connects agents like Claude, Cursor, Codex, or ChatGPT directly to the WhatsApp Business platform. What previously required jumping between the developer console, Business Manager, API references, and the editor is now requested via chat: creating the account, adding and verifying the phone number, registering the business for Cloud API access, preparing templates, testing, and checking Terms of Service requirements. For your company: this is the rare case where an announcement from a big platform translates into consulting hours you stop paying for, because the bottleneck for automating WhatsApp in Spain was never the idea, it was the setup. Two warnings before you jump in. First, that the agent configures does not mean the agent decides: message templates still go through Meta's approval and whoever responds to a customer remains responsible for what is said. Second, permission: if you give an agent the keys to your business account, give them the minimum role and review them when finished, which is exactly the lesson from the previous news item. Source
SK Hynix negotiates manufacturing memory in the US with Intel, and that ends up in the price of your next server
SK Hynix and Intel are in talks for the South Korean manufacturer to produce memory on US soil for the first time: the two scenarios on the table are for SK Hynix to occupy part of Intel's future Ohio plant or for them to set up a joint venture with major cloud providers, who are the ones trying to secure supply. Both companies emphasize that the talks are exploratory and nothing has been decided; SK Hynix is already building a $3.8 billion packaging plant in Indiana and so far does not manufacture wafers in the United States. The core of the matter is the memory shortage caused by AI demand. For your company: this is not distant geopolitics, it's your renewal budget. When hyperscalers fight for memory, the price of RAM for laptops, servers, and the cloud instances you rent by the hour goes up. If you have equipment or server renewals planned for 2027, ask for a price now with written validity, and in the cloud, check if it pays to commit to reserved capacity instead of going on-demand. It's the kind of boring decision that is either made six months in advance or not at all. Source
85 million in four months for a twenty-person company that fixes legacy software for big firms
Hang Ten Systems, founded in May by Vishal Sikka, former CEO of Infosys, has expanded its seed round with $53 million five weeks after closing the initial $32 million: $85 million in total, with Xora—Temasek's early-stage platform—leading this expansion and Mayfield the first, plus Aramco Ventures, Intel CEO Lip-Bu Tan, Micron CEO Sanjay Mehrotra, and Jerry Yang. It sells on its own framework, Hobie, which packages reusable AI skills for regulated sectors, and already has 21 major clients—Fresenius Kabi, Saudi Aramco, Siemens Energy—with several seven-figure contracts, all with a staff of between twenty and twenty-five people. Sikka's quote is the thesis: software development has shifted from writing code to defining requirements and validating system performance. For your company, there are two takeaways. First, the big money is going into modernizing what already exists, not new applications: if you're dragging along a custom ERP, Excel macros that sustain a process, or an application only one person understands, that work just got cheaper and it's time to ask for a quote again. Second, the step: twenty people billing seven-figure contracts to Siemens Energy is the operational definition of AI First, and what is outsourced there is not judgment, but the typing. Start by documenting on one page what your most fragile process actually does; without those written requirements, no AI is going to modernize it for you. Source
What to watch tomorrow?
If the AEPD expands the information on the incident—especially if they say what type of application was taken down—and if more notifications of the same profile start arriving, which is what would turn a case into a trend. And on the supply side, if the talks between SK Hynix and Intel result in an announcement with figures: that's where we'll see how much memory will cost in your next renewal.