noticias ia para empresas ciberseguridad openai

Money is flowing to AI that doesn't chat but decides: TypeSafe, the company behind Jev, has raised about $870 million at a $7.5 billion valuation, less than four weeks after launching it. AI now finds bugs faster than they get fixed: Anthropic is offering its scanner free to open-source projects, and Spain's banks are finalizing their Mythos plan for the ECB. And OpenAI says its revenue is running at about $50 billion a year, not $68 billion, while in Spain Autentic AI has raised €1.5 million to deliver market research in days. Picks up from Friday's radar, October 9.

Jev, the AI that doesn't write but decides, is now worth $7.5 billion: only let the machine decide above a threshold

TypeSafe AI, co-founded by Diogo Almeida, a former OpenAI researcher who worked on the methods behind ChatGPT, has raised about $870 million at a $7.5 billion valuation in a round led by Andreessen Horowitz, with Sequoia, according to Bloomberg. It came out of stealth on September 15 with a $40 million seed round and Jev, a model that doesn't generate text: it takes in an application's data and returns a fixed-format decision (an option, a score, a probability) along with its confidence level, to classify, route or score inside software. According to the company, it responds in 70 to 500 milliseconds, costs $0.042 per million input tokens (output is free), and is already used by a third of the Fortune 500. Its claim of 193 times faster and 444 times cheaper than a language model comes from four of its own test workflows, which the company itself puts at the high end of what to expect. For your business: almost everything you automate isn't writing, it's deciding. Does this invoice go to payment or to review? Is this ticket urgent? Do we approve this return? What matters there, whether you use Jev or any other model, is the confidence score: set a threshold (say, it decides on its own above 95%) and send everything else to a person. Every month, check how many automatic decisions went wrong, and adjust. And if the decision affects a person, have someone review it: Anthropic's usage policy, for one, already requires that. A process that decides on its own without a written threshold isn't UNATTENDED: it's running on luck. Source

Anthropic gives its bug scanner to open source for free and Spain's banks prepare their plan for the ECB: patching can't take months anymore

On Thursday, Anthropic launched OSS Scanner: free, recurring security reviews run with its most powerful models, Mythos included, for critical open-source projects that opt in. In six months, its models have found more than 29,000 potential vulnerabilities in widely used software, and it has only been able to review about 6,000 by hand; reports from the new service go out with no human review. In a test of 97 critical or high-severity bugs across 48 projects, 85 were valid, 11 more were real but duplicates, and only one was a false positive. Early testers include PostgreSQL, OpenSSL and curl, building blocks that sit under almost every website, and the same day it launched a program to protect power grids and factories with partners such as Rockwell, Hitachi, CrowdStrike and Deloitte. Meanwhile, Spain's banks are finalizing the Mythos plan the ECB wants by October 31: patch faster, inventory all their technology assets (third-party ones included), cut what's exposed to the internet and review their suppliers. For your business: bugs are now found by machines, and Anthropic admits that months often pass between finding one and fixing it; that gap is your risk. Ask whoever maintains your website or software, in writing, how many days it takes them to install a patch once it's released. If you sell to a bank, have that answer ready, because the ECB is asking banks to review their suppliers. And an AI-generated bug report is just an alert until someone reproduces it. It picks up where we left off with the annual pentest that's just a snapshot. If your patches depend on someone remembering, that part is stuck in MANUAL even if everything else runs on its own. Source

OpenAI's revenue is running at $50 billion a year, not $68 billion: before comparing vendors, ask how they count

OpenAI told investors its annualized revenue was about $50 billion at the end of September, not the $68 billion widely reported late that month, as the Financial Times first reported and CNBC confirmed. The higher figure included gross revenue from its partners, which is how Anthropic counts (it reported $65 billion at the end of July), and it let investors compare the two. In the third quarter, OpenAI's revenue run rate grew 77%, and the run rate of its enterprise business grew 107%. On Thursday, Nvidia shares fell 3%, Oracle nearly 6% and CoreWeave nearly 8%. For your business: two numbers for the same thing are $18 billion apart purely because of how they're counted. The same goes for what vendors promise you: "save 75%," "automate 80%." Before you sign, ask what's being counted (per token or per finished task, as we covered on October 6), compared with what, and over what period. The other takeaway: business customers are where OpenAI is growing fastest, so expect more sales pitches. It's a good time to negotiate price and contract length, not to sign blind. A vendor's number you haven't checked against your own data isn't data: it's marketing. Source

Autentic AI raises €1.5 million to turn 12-week market studies into days: your company can now ask customers every month

Spanish startup Autentic AI has closed a €1.5 million pre-seed round from Baobab Ventures (Carles Reina's fund), Lanai Ventures and Acurio Ventures. Its platform uses AI to interview hundreds of consumers or experts at once (real, verified people who get paid) and delivers results in days, with a researcher on every study. According to its founders, a traditional study takes about 12 weeks and costs tens or hundreds of thousands of euros. It already runs 1,000 interviews a day and aims to finish 2026 above 400,000, focused on consumer goods, retail, pharma and cosmetics, and consulting. For your business: testing a price, packaging, or a message before launching it is no longer just for big companies. But always ask the question Autentic itself pitches as its edge: who's behind each answer. A study built on AI-generated synthetic consumers, or on a panel full of bots, tells you what the model thinks, not what your customers think. Ask to see how they verify participants, and start with a decision you were going to make on gut feel. Setting prices on gut feel once a year is MANUAL; asking real customers every month is a step up the ladder. Source

What to watch tomorrow

Whether any Jev customer goes on the record to back TypeSafe's numbers, and how many Spanish banks file their Mythos plan before the 31st. And on Tuesday, October 13, TechCrunch Disrupt kicks off in San Francisco, running through the 15th.