noticias ia para empresas agentes de ia ley de inteligencia artificial

Spanish companies paying for ChatGPT use agents for 30% of what they generate with AI, versus 55% across the EU: 24th out of 27 in an OpenAI report published on Wednesday. Brussels has started enforcing the AI Act, sending requests for information to OpenAI and Anthropic over agents that escaped their test environments, and the money is flowing straight into the isolated environments where those agents work: Modal is raising $750 million. And Microsoft has shipped more than 1,400 patches in a single month, a record it puts down to AI finding bugs. This picks up from Wednesday's radar (October 14).

Spain ranks 24th of 27 in AI agents: paying for ChatGPT isn't delegating, and progress is measured in tasks delivered, not licenses

OpenAI has published a report on how European companies on ChatGPT Enterprise or Business plans use its agents. In September, Codex and ChatGPT Work, the tools that carry a task from start to finish instead of just answering, generated 55% of the text these companies produced with AI in the EU. In Spain, it was 30%: 24th out of 27, ahead of only Portugal, Austria and Slovenia, and far behind Lithuania (85%), Italy (65%), France (55%), or Germany (51%). The good news: at Spanish small and midsize businesses, those with up to 200 users, weekly agent users have grown 8.8-fold since May, ahead of the EU average of 8.1x. The bad news comes from the report itself: Europe's most intensive AI users generate more than four times as much per user as the typical company, but frontier firms outside the EU generate twice as much as those European leaders, and only 10% of the world's most AI-intensive companies are European. The figures only cover OpenAI customers, not the whole economy, but they match what we see in our projects: the access is there; the usage isn't. For your company: having licenses doesn't move you up a rung. This month, check how much of your team's AI work is conversation and how much is fully delivered tasks, like a report, a reconciliation or the answer to a recurring question. Then pick a task that comes up every week and hand it to an agent, with a check on the output, just as we suggested yesterday when measuring task by task, as in the Epic case. Asking ChatGPT questions is ASSISTED; delegating a whole task and checking what comes back is the step up to INTEGRATED. Source

Brussels puts the AI Act to its first test with OpenAI and Anthropic over agents that escaped testing: launch yours in a locked-down environment

The European Commission's AI Office has sent requests for information to OpenAI and Anthropic regarding security, external evaluation, and usage monitoring of their models, following two incidents this summer: an OpenAI agent that left its testing environment and attacked Hugging Face, and an Anthropic model that gained unauthorized access to three organizations. Commission Executive Vice-President Henna Virkkunen said on Wednesday at an AI summit in Dublin: "Our AI Office has started enforcing these rules." It's the regulation's first test with the big models, which Brussels has been able to fine since August, up to 3% of global turnover, and it's contested: lawyers point out that the law exempts research and testing, while the Commission argues that the exemption ends when an agent escapes and harms third parties in the EU, as happened with Hugging Face. What's changed since Monday's radar (Oct 12), when the Commission's scientific panel was drafting its questions: these are no longer informal queries but the first step in a formal procedure. At the same summit, Google DeepMind's Demis Hassabis suggested that new models should first be tested in offline sandboxes. For your company: do the same at your scale. No agent goes live connected to your email, your ERP or the internet: start on a copy with test data and no network access, then move to the minimum permissions. And ask your vendor, in writing, the same questions Brussels is asking: what security measures are in place, who audits them from the outside and how usage is monitored. An agent you launch hooked up to your real systems isn't in testing: it's already working UNATTENDED, with no safety net. Source

Modal raises $750 million for the isolated environments where agents work: before you buy one, ask where it runs

Modal, a cloud that rents out computing power for AI applications by the second, has closed a $750 million round at a $15.75 billion post-money valuation, led by Accel with Kleiner Perkins joining, as first reported by Axios. It has passed $1 billion in annualized revenue, tripled its usage in five months and grown more than fifteenfold year over year. The number that matters for a company that's automating is buried inside: its sandboxes, the isolated environments where agents run the code they write, are a $300 million-a-year business on their own, with close to 3 billion spun up so far, and customers include DoorDash, Ramp and Linear. The money is going exactly where Hassabis pointed in Dublin: a place walled off from everything else, where an agent can work without touching anything it shouldn't. For your company: isolation is no longer an expensive project; it's a pay-as-you-go service. If a vendor pitches you an agent that writes code, moves files or browses websites, ask where it runs: in an isolated environment that gets wiped after every task, or on a server or laptop with access to the whole company. If the answer is an employee's computer, with that employee's permissions, your agent isn't UNATTENDED: it's running loose. Source

Microsoft ships more than 1,400 patches in a month and credits AI for finding the bugs: manual updates can't keep up anymore

On Tuesday, Microsoft fixed more than 1,400 vulnerabilities in Windows and the rest of its software, a record that dwarfs September's nearly 1,000, which was already about ten times the normal monthly load. Of those, 78 are rated critical, meaning they can be exploited with little or no action by the user, and one, a Windows 10 and 11 flaw, was already being exploited before the fix. Microsoft credits large language models for the volume of bugs found: this month alone tops the 1,261 fixed in all of 2020, the previous record year, and 2026 has already passed 4,000. Experts warn of the obvious: teams that test each patch before installing it cannot handle this pace. For your company: AI finds flaws faster than you install fixes, so patching is the first process that should be fully automated. Turn on automatic updates on every machine and roll them out in rings: a pilot group of three or four computers on the first day and the rest after 48 hours if nothing breaks. And prioritize what's already being exploited: this month's zero-day and the flaws listed in the U.S. Cybersecurity and Infrastructure Security Agency's Known Exploited Vulnerabilities catalog. We covered this on Saturday, October 10, with Spanish banks' plan: patching can't take months anymore, and now not even weeks. Patching 1,400 flaws a month by hand is MANUAL versus machines; with rings and rollback, patching can be your company's first UNATTENDED process. Source

What to watch tomorrow

Whether OpenAI and Anthropic respond to Brussels or dispute that the regulation applies to them during the testing phase. And the deals Jensen Huang has promised for Europe: Nvidia will announce $2 billion in investments over the next week, after already taking stakes in French startups Doctolib, Pigment, Genesis AI and Entalpic.