noticias radar ia ai act seguridad automatizacion

The radar from yesterday was about who owns your tools; today is about the trail they leave. Anthropic has detailed how it marks everything Claude writes—an invisible signal that survives light editing, with no option to disable it—and Google has done just the opposite with images: removing the visible stamp and leaving only the invisible one. In parallel, Anthropic's acquisition of Decart rises to approximately $7 billion and could be signed in weeks, ahead of its IPO.

Anthropic to pay around $7 billion for the company that makes running models cheaper

The deal for Decart, the Israeli company whose technology allows models to run up to eight times faster on the same hardware, has risen from the $6 billion discussed on Thursday to around $7 billion, paid mostly in Anthropic stock; advanced drafts are already being exchanged and it could be signed in weeks, before the IPO planned for September or October. The detail that tells the story is who lost out: Nvidia, which was already an investor in Decart, put a higher valuation on the table, and yet the founders and Sequoia—its largest shareholder—preferred Anthropic. Decart is three years old, has about a hundred employees, and has raised 450 million. For your company: translated to your bill, your model provider is buying efficiency, not capacity. When a lab spends 7 billion to better squeeze the chips it already has, it means the margin for the next stage comes from lowering the cost per token, not from raising your price—but that saving takes months to reach the rates, and it reaches new models first, not the one you have locked in production. The practical takeaway: don't anchor your automations to a specific model, and review every quarter if the same work is already being done cheaper in the next version. This isn't stock advice; it's one of your providers reordering its cost structure right before having shareholders. Source

Everything Claude writes for you carries an invisible mark, and now we know how much it withstands

Anthropic published the details of its watermark on August 15: all models released since August 2 embed a statistical signal in the text—choosing between equivalent words—and it is applied in the API, in claude.ai, in Claude Code, and in models served from AWS, Google Cloud, and Microsoft Foundry. The mark survives light editing and dilutes the more you rewrite: it only disappears if you change practically every word. Generated code carries almost no mark because there is nowhere to hide it, although it may appear in comments. Anthropic uses Google DeepMind's SynthID-Text method, says the cost in speed and tokens is negligible, is preparing a public detection API, and offers no way to deactivate it. The engine behind all this is Article 50 of the European AI Act, enforceable since August 2. For your company: assume that any text coming out of a model is detectable as such, and that soon anyone will be able to check it with an API call. That is not a problem if you have planned for it, but it is if someone on your team delivers something to a client, a court, or a university that they claimed to have written by hand. This week, write a two-line rule: in which deliverables generated text is allowed, who reviews it, and if it is declared. And if you sell content, put it in your contract before the client asks you. Source

Google removes the visible stamp from its generated images and leaves only the unseen one

The same weekend, Google announced that the visible watermark on its generations is becoming optional: it can be deactivated from Settings in the Gemini app, in the Flow video editor, and soon in Search, for the Nano Banana, Omni, and Lyria models. What cannot be deactivated is the invisible SynthID mark nor the C2PA metadata, which continue to travel with the file. Google has also released Credentio, an open-source library so that any developer can validate C2PA credentials in their own application. For your company: the daily consequence is that you can no longer trust what you see. The fact that an image does not have a label means nothing, and from now on, the only serious check is reading the metadata—something you can already set up yourself with that library, for free. Two specific uses: if you receive graphic documentation from providers, candidates, or claims, add an automatic C2PA check before it reaches a human; and if you publish generated material, leave the visible stamp on when the context calls for it, because removing it is now your decision and not the provider's. Source

Your team's AI accounts don't close like others, and that's an exit hole

A guide published on August 15 reviews how to check if someone has accessed your AI accounts and reveals an uncomfortable asymmetry between platforms: in ChatGPT, there is a list of active sessions in Settings and the password can be changed; in Claude, there is a list of sessions and a mass logout, but there is no password to change because entry is via email; and in Perplexity, you can't even see which sessions are open, only close them all and re-enter with a code sent by email. For your company: if the door to a tool is email, whoever controls the inbox controls the tool—and that changes your employee offboarding process. When someone leaves, cutting off email isn't enough: you must close all sessions on each AI platform, one by one, because an open session on a personal laptop remains live. Today, make a list of the AI tools your team uses, who the account holder is for each, and which email opens it; add closing sessions to the same script you already use to revoke CRM access. And activate two-factor authentication where it exists, which in several of them is the only thing there is. Source

What to watch this week?

DeepSeek's new rates went into effect yesterday at 16:00 UTC, so you'll see the first bill with the actual price this week. And on Saturday the 23rd, generated data from December 29th will be deleted in Manus: if anyone on your team uses it, the backup is due before Friday.

Watch the 1-minute video