noticias ia para empresas ciberseguridad agentes de ia

OpenAI built 948 working Linux exploits at $1,279 apiece, and that shortens the time you have to patch. Two warnings hit everyday office tools: a browser extension can get Claude to run code on your computer, and Copilot Chat now has caps Microsoft won't publish. The money is moving into the back office: Kabilio raised €10 million to bring agents to accounting firms, and Dropbox is paying $350 million for Tigris. This follows Thursday's radar.

OpenAI built 948 Linux exploits at $1,279 each: shorten your patching window

OpenAI released GPT-6 Cyber on Thursday with the highest risk rating on its scale, "Critical," and for now only to a small group of approved customers in its Daybreak Red program. A day earlier it detailed a run from mid-October: in 30 hours, its agents reproduced 1,724 Linux kernel flaws and built 948 working exploits, 588 of them on previously unknown bugs, for $1,213,430. These don't break in from outside. They turn an unprivileged account into an administrator. The model also found 81 new high-severity flaws in Firefox, which Mozilla is now fixing. The numbers are OpenAI's own, and it's reporting the Linux findings with proposed patches. Google, for its part, says monthly vulnerability disclosures have doubled, from 5,045 in January to 10,740 in August. And Acronis just described a single person who targeted hundreds of organizations in 11 countries with 12 agents, rented GPUs and weaker local models, because commercial ones refused. For your company: you don't need this model and you can't buy it. What changes is the clock. Write down today what's exposed to the internet, who patches it and how many days it takes. Ask your hosting or IT provider how long a critical Linux patch takes, and get it in writing. Since these flaws start from a low-level account, delete the ones nobody uses and turn on two-factor authentication for the rest. Source, Source and Source

A browser extension can get Claude to run code on your computer: keep personal and work accounts apart

Researcher Gal Weizman of Forever Security told The Register that any extension with access to claude.ai can use Dispatch, the feature for sending tasks remotely, and tell Claude Desktop to run code on that computer or on another one signed in to the same account. Anthropic accepted the report as valid in May, closed it as a duplicate of an issue it was already tracking, and told him it planned to retire Dispatch at the end of October. On October 30 it was still available to existing users, and the company didn't answer The Register's questions. Weizman argues this is a design problem that lives on in Claude Cowork and Claude Code Remote Control. The attack needs a malicious or hijacked extension and an open session. His demo only opens the calculator, and the article doesn't cite real attacks. The number to worry about is a different one: at six customers of his firm, which sells security, personal accounts made up between 18% and 71% of the Claude accounts seen inside the company, and a shared account is the bridge from a home computer to a work laptop. For your company: if you use Claude, hand out company accounts and ban personal ones on work machines. Keep a closed list of approved extensions, which Chrome and Edge let you enforce by policy, open Claude in a browser profile with no extensions, and turn off Dispatch if nobody uses it. It's the same logic as CrowdStrike's kill switch: an agent with access to your computer needs a control that doesn't depend on the agent. Source

Copilot Chat now has caps Microsoft won't publish, and Apple is rationing AI for its engineers: set the budget yourself

Since November 2, at organizations with up to 2,000 seats, Microsoft has been applying "fair use" limits to more Copilot Chat features for anyone without a paid Microsoft 365 Copilot license. Larger organizations follow on November 9. It doesn't publish the numbers because they're "subject to change," they vary during the day, and admins can't see or adjust them. Users who hit a cap get a notice saying when to retry and, at most organizations, an option to request a license from their admin, or to buy one if the company allows it. Separately, Bloomberg reported on Wednesday that Apple cut its engineers' daily AI spending cap from $300 to $200, with cuts to as low as $100 in the works, and that it's weighing $25 outside engineering. For your company: at up to 2,000 seats, you're already in. Tell whoever runs support, turn off direct purchases in the admin center if you don't want surprise licenses, and spend two weeks logging who hits the cap and doing what. That list tells you who's worth a license and who isn't. Don't build a critical process on a limit you can't see. And set an AI budget per person per month: if Apple rations it, your bill won't control itself either. Source, Source and Source

Kabilio raises €10 million to bring agents to nearly 500 accounting firms: ask yours who signs off

Barcelona-based Kabilio closed a €10 million round led by Picus Capital and Visionaries, with Spain's CDTI taking part, for €14 million in total. In a year it has grown from about 100 accounting firms to nearly 500, and it claims to save them around 60% of their bookkeeping and reconciliation time, a figure that comes from the company itself. The money pays for an agent, entering beta this month, that gathers invoices and bank statements from inboxes, reconciles accounts, checks entries and chases clients for what's missing by email or WhatsApp. The firm decides, client by client, how much it does on its own. It isn't alone: Belgium's Ravical raised $17 million last week. For your company: your first agent may not be one you buy. It may arrive through your accountant. Ask which tool they use with your data and whether it's listed as a sub-processor, who reviews and signs off each entry, and how you'll know that a WhatsApp asking for documents or a change of bank account really comes from them. If you run an accounting firm, test that 60% on five clients before you buy, and start with approval on everything: granting autonomy client by client is the sensible way to climb to the UNATTENDED rung of the AI First ladder. Source

Dropbox pays $350 million for Tigris: file storage gets ready for agents

Dropbox announced on Thursday that it's buying Tigris Data, a cloud storage specialist, for $350 million, apparently its largest acquisition, with closing expected in the fourth quarter. The deal will support DBX Agent Cloud, storage for people building models and agents that works with tools built for Amazon S3. Customers will also be able to ask AI to do tasks with their files, choose a model and connect Dropbox to ChatGPT, Claude and Muse. That starts in the US, the UK and Australia, with broader availability planned for 2027. It fits the pattern of OpenAI's Pierre deal: storing what agents use now commands a high price. For your company: in Spain there's no rush, but there is homework. Before you connect your folders to any agent, check who can access what, clean up anything shared with the whole company, and move payroll and contracts out of it. Source

What to watch tomorrow

On November 9 the Copilot Chat caps reach larger organizations. Also watch whether Anthropic responds on Dispatch and when OpenAI widens access to GPT-6 Cyber.