noticias radar ia agentes de ia ciberseguridad mantenimiento predictivo

Yesterday's radar was about who can touch what you have connected to AI; today's is about the gap between what is bought and what actually works. A report published yesterday puts a figure on it: 99% of companies plan to deploy agents and only between 9% and 14% have them in production. Meanwhile, money is going toward what is already generating revenue: Munich Re pays $575 million for a cyber-risk insurer that monitors the internal systems of the SMEs it insures, Google secures its silicon with a purchase right for $12.18 billion in Marvell shares, and the less flashy side of automation—predictive maintenance—launches a sensor that lasts ten years on one battery.

99% of companies want agents; nine out of ten get stuck in the pilot phase

A report by Ness Digital Engineering released yesterday places 99% of companies planning to deploy agentic AI and only between 9% and 14% with something in production—a stretch they call the "valley of death": the distance between the proof of concept that looks good in the demo and the process that runs every day. The causes it points out are not model-related, but house-related: distrust of a probabilistic system that doesn't always respond the same way, lack of visible improvement for the person using the tool, and lack of preparation of data, APIs, and the processes themselves before scaling. For your company: the useful data isn't the 99%, it's that the filter separating the two groups is applied before hiring anything. Three concrete things. One, choose a process with volume and a success criterion that can be measured in a number—invoices matched per day, time to first response, orders fulfilled without being touched—and record that number today, because without a baseline, there's no way to know if the agent is useful. Two, demand that the agent write in the system where your people already work: if the result ends up in a chat that someone has to copy and paste into the ERP, you've added a step instead of removing one, and that's where most pilots die. Three, set a decision date: six weeks and a meeting where it's decided whether it goes to production or is shut down. If you're interested in the complete order, you have it in our 90-day roadmap. Source

Your cyber insurance stops being a yearly piece of paper: Munich Re pays 575 million for the one that looks inside you

Munich Re announced yesterday the acquisition of At-Bay for an enterprise value of $575 million, its largest declared entry into direct cyber insurance, with closing expected in the first quarter of 2027 and fitting within Hartford Steam Boiler, its specialized subsidiary. The price is equivalent to about twice At-Bay's $278 million in gross written premiums and falls well below the $1.35 billion valuation from its 2021 round. What it's buying isn't a portfolio of policies: it's InsurSec, a platform that monitors the insured's systems throughout the life of the contract and feeds real-time security data into underwriting—the company claims this covers risk factors present in 86% of its clients' claims. At-Bay targets SMEs, where cyber insurance penetration is between 10% and 20% compared to 60-70% for large companies, and where 43% of breaches originate. For your company: this anticipates how your policy will be renewed. If your broker renews it without asking for a single piece of technical evidence, you're in the old model, and the new model sets the premium based on what it sees: multi-factor authentication, truly restored backups, equipment control and, very soon, the inventory of which AI agents have access to what. There is a question worth asking in writing before the next renewal that almost no policy clearly answers today: is an incident caused by an in-house AI covered—an agent sending customer data to the wrong recipient, an assistant with manipulated memory, an error by a model provider—or does it fall under the unauthorized act exclusion? Get them to answer in writing. Source

Google secures its own chips with $12.18 billion in shares it hasn't paid for

Marvell granted Google yesterday a purchase right on 58,970,907 of its shares at $206.58 each—$12.18 billion if fully exercised, which would make Google its fifth-largest shareholder—and the mechanism is the interesting part: about 1.4 million shares vest in the first year and the rest are unlocked for every $500 million in chip purchases through fiscal year 2033, a path that could result in about $120 billion in custom silicon sales for Marvell. The deal covers what surrounds Google's TPU—inference accelerators, storage, and network controllers—Marvell's stock rose as much as 14% and Broadcom, the historical provider for that task, fell nearly 5%; Citizens estimates that TPU infrastructure contributes about $3 billion to Alphabet this year and $25 billion in 2027. For your company: there's a market reading and one you can use on Monday. The market reading: the alternative to Nvidia for running models stops being a promise and moves to having a signed supply chain until 2033, so inference capacity will be less concentrated than it is today; at your scale, that translates into not tying your automation to a single provider—an abstraction layer upfront and the same battery of tests passed to two models, to be able to switch without rewriting. The other is pure negotiation, because this agreement is a volume contract paid in shares: when your software provider offers you a discount in exchange for a multi-year commitment, the counterpart is requested just like here, in tiers and with an exit—fixed price per tier, portability of your data in a usable format, and cancellation without penalty if service is not met. Source

The AI that actually pays off in a factory doesn't talk: it listens to a bearing

TDK presented yesterday the edgeRX Pro, a sensor node that combines a six-axis inertial unit, magnetometer, digital microphone, and temperature sensor, and runs models on the device itself instead of sending the signal to the cloud: IP67 housing for harsh environments, up to ten years of battery life, and USB power when wiring is possible. The cases it lists are the usual suspects and that's why they matter—compressed air leaks, acoustic anomalies, misalignment, and bearings—in manufacturing, energy, HVAC, and warehouses. For your company: predictive maintenance is the automation with the easiest ROI to demonstrate and the one that appears least in presentations because it has no conversation or pretty screen. Having the model run inside the sensor takes away the two common excuses for not starting: there's no need to upload continuous vibration anywhere or ask systems to open anything. The order is the same as in today's first news item and for the same reason: choose the five machines whose downtime actually costs you money, record for a month how many hours you stop without planning and how much each hour costs you, and only then install. If you can't write down the savings as avoided downtime hours, what you've bought is a dashboard. Source

What to watch for tomorrow?

What Alation ends up reporting, after confirming yesterday unauthorized access to one of its systems without yet saying what was taken: it's the data catalog for over 500 large companies, approximately half of the Fortune 1000, and a catalog doesn't store your data but it does store the map of where it is and who uses it. The question it leaves for the rest is not whether it affects you as their client, but whether your providers would notify you in the same timeframe. And if Broadcom makes a move after losing part of Google's business.

Watch the 1-minute video