noticias radar ia coste de la ia seguridad ia agentes de ia

Yesterday's radar was about having your AI at home instead of renting it; today the conversation moves to cost and what you leave behind. Nvidia has released a free router that looks at each task and sends it to the cheapest model capable of solving it—in their tests, at a third of the cost without losing accuracy—along with a small model for agents that runs on a normal computer. And on the uncomfortable side: academic work demonstrates that the "encrypted reasoning" blocks returned by OpenAI, Anthropic, and Google can be decrypted, and real passwords have already been extracted from public repositories.

You no longer choose a model: a free router sends each task to the cheapest one that can do it

Nvidia published NeMo Switchyard on August 11, an open-source library that sits in front of your models and decides, request by request, which one is the most capable and efficient for that specific job, without rewriting the application; in its own tests, it maintains top-tier accuracy while spending about a third of what it would cost to solve everything with Opus 4.8. On the same day, it released Nemotron 3.5 Lightning, an open model with 30 billion parameters, of which only 3 billion are activated per token (permissive OpenMDW-1.1 license): up to four times faster in generation than others of its size, it completes ten thousand agent tasks 30% faster and runs on an RTX 5090, a DGX Spark, or even a Jetson with LM Studio, llama.cpp, or Ollama; it can be downloaded from Hugging Face, ModelScope, OpenRouter, and build.nvidia.com. For your company: the cost of an automation is no longer decided by which model you chose at the beginning, but by which model each step deserves—classifying an email isn't worth the same as drafting a proposal. Start by measuring what percentage of your calls are routine tasks (classifying, extracting data, summarizing): that percentage is your immediate savings as soon as you route it to a small model and leave the expensive one only for the hard stuff. Source

The encrypted reasoning blocks you keep in your logs are not a safe: passwords have already been pulled from them

Work published on August 10 demonstrates that those encrypted reasoning blocks that major providers return to the client to be resent in each turn are interchangeable between sessions, between users, and between models from the same provider: it is enough to inject a powerful model's block into a weaker and less protected model from the same provider for it to spit it out in plain text, without touching the good model. It works on Anthropic, OpenAI, and Google, and by decrypting 315,320 blocks collected from public repositories, the authors recovered 367 pieces of personal data and 182 credentials; the same blocks also serve to sneak invisible instructions to the model. For your company: if your application or your team keeps session logs with those blocks inside—in a repository, in a support ticket, in a shared folder—treat them as clear text with sensitive data, because that's what they are. Today: remove them from any public repository, remove them from the logs you share with third parties, and distrust any reasoning block that comes to you from the outside. Source

Spotify starts labeling AI-made identities and leaving them out of its recommendations

Spotify announced on August 11 the "AI Persona" label, which warns that an artist's identity—name and image—is generated with AI and does not correspond to a real person. Since that same day, artists can self-declare through Spotify for Artists; labels will start appearing in mid-September on profiles, in search, and in playlists, and Spotify will also apply them on its own after reviewing profiles with photorealistic AI-generated images, notifying the affected party and allowing for an appeal. The important part is the consequence: by default, this content is left out of personalized recommendations unless the listener actively follows the profile. For your company: platforms are moving from "declare it's AI"—which the European regulation Article 50 has already required since August 2—to "declare it's AI and I will distribute you less"; if part of your marketing relies on synthetic characters, expect a label and less organic reach. Declaring it yourself is cheaper than having them tag it for you, and AI still comes for free where it doesn't penalize: in the production behind a real identity. Source

If anyone on your team uses Manus, have them make a backup before the 23rd

Manus announced on August 11 that it is operating as an independent company again after its acquisition by Meta was undone, a deal Beijing ordered reversed in April as part of its control over US investment in Chinese frontier tech startups. The fine print is what affects the user: data generated by them since December 29, 2025, will be deleted at the end of August—a backup must be made before the 23rd to be able to restore it later—with notifications via the app and email, and the company says it will not charge during the transition and will compensate those affected. For your company: this is yesterday's warning about depending on a single provider, but concrete and with a date on the calendar: a regulatory decision ten thousand kilometers away can erase your work history with two weeks' notice. Today: if anyone on your team uses Manus, have them export; and as a general rule, what you generate within an AI tool (prompts, agent history, results) should have a backup on your own systems, not just theirs. Source

What to watch for tomorrow?

If providers react to the encrypted block flaw—rotating them, tying them to a single session—and if ready-to-use routers appear outside the Nvidia world, which is where this turns into real savings for an SME. And a question for today: of everything your company has generated within AI tools this year, how much survives if your account is closed tomorrow?

Watch the 1-minute video