noticias ia para empresas ciberseguridad agentes de ia

AI agents are already attacking on their own: one broke into DIVD, a Dutch institute of ethical hackers, through two unknown flaws in its help desk software and took over the server in seconds, while Epic, the medical records giant, has paused most of its development to fix what another AI found. The money is following: Armadin, from the founder of Mandiant, raises $255.5 million to attack its clients with agent swarms before the bad guys do, and Wall Street is already looking for more people to oversee agents than to build models. And Brussels is preparing the gatekeeper label for AWS and Azure for November. Picks up from the Friday, October 2 radar.

An AI agent broke into DIVD, the Dutch bug hunters, through its help desk software and got root in seconds: check what you have exposed

The Dutch Institute for Vulnerability Disclosure (DIVD), a non-profit organization of volunteers who hunt for security flaws, has laid out, step by step, how it was attacked: on September 21, someone got in through two previously unknown flaws in Zammad, the open-source software it used to manage support tickets, and by chaining them went from hijacking a session to full control of the machine (root) "in seconds." DIVD maintains that the attacker was an AI agent: the attack was "loud and very messy," the agent decided each next step itself, and it left notes in its scripts explaining why what it was doing was okay and "really not phishing." On Thursday, it confirmed that volunteer data had been taken, such as DIVD email addresses, and that network segmentation stopped the attacker from getting any further; it spotted the attack on the 22nd and had notified the Dutch data protection authority by the 24th. The flaws (CVE-2026-102489 and CVE-2026-102490) score 9.4 out of 10 when chained, and DIVD advises all Zammad users to upgrade to version 7 or take it offline. On the defense side, Epic, whose MyChart holds more than 320 million patient records in the US, has paused most of its product development—about six weeks, according to its founder—to fix flaws found by Mythos, Anthropic's cybersecurity model: some customer configurations let outsiders view records without leaving a trace in the logs, its chief security officer told The New York Times. For your company: if you run Zammad, this is today's job. If you don't, the lesson still applies: whatever you have open to the internet—your help desk, your web forms, your web-based CRM or ERP, your WordPress site and its plugins—is no longer probed by a human attacker now and then, but by an agent that never stops. Three things: a list of everything exposed, with an owner who patches each piece and a deadline in days, not months; what's exposed walled off from what matters, as DIVD had it, so one door doesn't open the whole house; and a written answer to who notifies the data protection authority (in Spain, the AEPD), because the GDPR gives you 72 hours. Ask your critical software vendors how long they take to ship a patch. As with the agents from the September 30 flash, you can't defend what isn't on the list. Only with someone to patch it can a tool support an UNATTENDED process. Source

Armadin, from the founder of Mandiant, raises $255.5 million to attack its clients with agent swarms: the annual pentest is no longer enough

Kevin Mandia, who founded Mandiant and sold it to Google for $5.4 billion in 2022, has closed a $255.5 million Series B round for his new company, Armadin, at a valuation of over $2.5 billion, led by Andreessen Horowitz and Accel. This comes six months after a $190 million Series A: over $445 million raised in total. It sells penetration testing taken to the extreme: instead of specialists who try to break in once a year and hand over a report, swarms of agents that attack its clients nonstop, chaining flaws together, to close the holes before others use them—including rogue agents from the AI labs themselves, like the OpenAI one that escaped its test environment (September 27 flash). For your company: money is flowing into automated defense because the attack is already automated, and it's a category that tends to end up with the giants—Mandiant ended up at Google. A small business won't hire Armadin, but it can change the question it asks its IT or security provider: not "when is the pentest due?", but "how often do you check what I have open to the internet and in how many days do you notify me of a flaw?". A yearly report is a snapshot; the DIVD attacker only needed seconds. And if you sign a multi-year security contract, get in writing what happens to your contract and your data if the provider changes hands. Source

The fastest-growing skill on Wall Street is coordinating agents (+1,721%), and banks now want nearly twice as many people to oversee them as to build them

AI-related job openings at banks like JPMorgan Chase, Citigroup, and Capital One have risen 49% this year compared to 2025, reaching 139,819, according to an analysis by hiring data firm Draup for CNBC. The fastest-growing area is "agent orchestration," the ability to design several agents that work together on a single task: mentions are up 1,721%. And overseeing now outweighs building: governance, risk, and control skills total more than 16,000 mentions, almost double the 8,400 for training, deploying, and operating models, and "responsible AI" mentions are up 657%. The example from Vijay Swaminathan, CEO of Draup, applies to any office: setting up agents to approve vacation requests uncovers a web of edge cases and exceptions. For your company: the bottleneck isn't the model; it's the person who knows which agents are needed, what each one does, and when a human needs to step in. Banks are solving this by training people who already know the business; in an SME, it's the same: before buying agents, get someone who knows the process—they don't have to be from IT—to write down its exceptions; if vacations have them, your orders and invoices do too. It's the "never decides alone" list, applied to every process. Only with that list can a process move from INTEGRATED to UNATTENDED. Source

Brussels readies gatekeeper status for AWS and Azure for November: switching clouds should get cheaper in 2027

The European Commission is finalizing the decision that will designate Amazon Web Services and Microsoft Azure, the two largest clouds in the EU, as "gatekeepers" under the Digital Markets Act (DMA), according to Bloomberg; it is planned for November, although it remains a draft and the date could move. In June, the Commission already shared its preliminary view: even though they fall short of the law's thresholds, they lock customers in with high switching costs, and their AI tools and alliances have become a "decisive factor" in cloud contracting. Now it says no final decision has been taken. If confirmed, they will have six months to comply: interoperability, easier data portability, and a ban on favoring their own services, with fines of up to 10% of their global turnover. For your company: this adds to what is already law. From January 12, 2027, the EU Data Act bans charging customers for switching cloud providers (Article 29), and that covers more than AWS and Azure: it also applies to the business software you pay for by subscription. If you are negotiating a spending commitment on Azure or AWS now, or renewing a cloud tool, don't sign for three years without an exit clause: the switch is what's free, not the commitment. And get it in writing: in what format, and how fast, they'll hand back your data and workflows. An automation tied to a single cloud can run UNATTENDED, but it isn't fully yours until you know how to get it out. Source

What to watch tomorrow?

Whether Zammad confirms the fix for the second flaw, the one that gives root access and affects all its versions, and whether more victims turn up in the case DIVD opened to notify them. And whether Amazon or Microsoft responds to the Bloomberg exclusive before Brussels decides.