noticias radar ia agentes de ia seguridad ia openai

Yesterday's radar was about what happens when AI fails; today's is about what happens when it gets too good. OpenAI is keeping its largest frontier training on pause because its internal model may have crossed the offensive capability threshold that its own preparedness framework calls critical, and it has imposed monitoring that consumes around twenty percent of the compute it supervises; on the same day, Anthropic has allowed Claude to send emails from your Gmail without asking permission every time. And in money, capital is paying for plumbing and paperwork: Temporal is negotiating about $500 million at a valuation of more than $12 billion, and Rillet is raising $100 million at $1 billion.

OpenAI has stopped its largest training because its own model is too good at attacking

OpenAI published on August 18 that its largest frontier reinforcement training remains halted while it conducts smaller-scale tests to validate safeguards: the decision stems from August 7, when it concluded that its internal model Astra can reach the cyber capability threshold that its preparedness framework classifies as critical. Since then, it requires monitoring of all Astra inference with tools—not just training and evaluations—with classifiers that inspect the model's internal activity at every sampled token and escalate to automated researchers who review actions on tools; the alert must arrive in less than thirty minutes, and the surcharge is around twenty percent of the monitored inference compute. For your company: there are two takeaways. The first is that your provider just put in writing that a model with tools is an attack surface, and their response wasn't a legal clause but looking at every token and accepting a twenty percent higher cost; copy this criteria to your scale, because the agent you've connected to your email, your ERP, or your repository needs logs of every action, expiring credentials, a domain whitelist, and above all, someone to read that log. The second is about schedule and price: this morning's flash edition mentioned that no new compute capacity is coming until 2028, and now you know that one-fifth of what exists is going toward watching what the other four-fifths do. If your plan for next year assumes agents that are more powerful, faster, and cheaper all at once, this announcement says that security is going to take its cut first. Source

Claude can now click send in your Gmail, and your administrator decides that

Anthropic enabled on August 18 for Claude to reply to, send, and forward emails within Gmail through its Google Workspace connector, available on all paid plans. By default, it asks for approval before each action, but the user can remove that approval and let it send on its own; in Team and Enterprise plans, account owners are the ones who decide if members can activate that mode. For your company: the checkbox that separates a draft from a sent email will be marked by each employee unless you decide beforehand, so that policy should be written this week and not on the day of an incident. It is reasonable to allow sending without confirmation only in specific, low-risk mailboxes—notifications, confirmations, internal reminders—and keep it closed in the mailbox used for signing quotes or responding to complaints. And it's worth being clear before signing off: an email sent by an agent goes out with your domain and your signature, so to the client receiving it, and to the GDPR if it contains third-party data, you sent it. Source

Today's big money isn't going to a model: it's going toward ensuring your automation doesn't stop halfway

Temporal is negotiating a round of about $500 million at a valuation of at least $12 billion before new money, according to Bloomberg: more than double the $5 billion it closed at in February when it raised $300 million. What it sells is not a model or an agent, but an open-source orchestration platform that ensures a multi-step process survives its own failures—if step four of nine fails, the process resumes where it was instead of starting from scratch or hanging in the middle. For your company: that capital is paying for this plumbing, and not another model, is the industry's diagnosis of why AI automations break. They don't break because the model writes poorly; they break because one step failed and left the order half-processed, the customer half-onboarded, and no one the wiser. Applying this doesn't cost a funding round: it requires each of your automations to be resumable—saving its state, ensuring repeating a step doesn't duplicate a charge or an email, and retrying automatically with exponential backoff—and asking whoever built it for you today exactly what happens when the model returns an error mid-flow. If the answer is that it's relaunched manually, that is your technical debt, and you already know what it's worth on the market. Source

They're going to sell you AI accounting this fall: the question that separates it from a demo

Rillet, a two-year-old company, closed a $100 million Series C on August 18 led by ICONIQ, valuing it at $1 billion: it's their third round in twelve months, exceeding $200 million raised in total, with Sequoia, Andreessen Horowitz, and Oak HC/FT returning while Bain Capital Ventures, Battery, FirstMark, Scale, and Creandum join. It sells an AI-native accounting ERP whose sales pitch is replacing legacy accounting software and giving weekends back to the CFO—meaning, automating the monthly close. For your company: this isn't about investing; it's about the fact that the category where you hire your ERP and your consultancy is being rebuilt with fresh cash, and by the end of the year, you'll be offered AI in both places. The filter to avoid buying vaporware is the close: how many days it takes you today to close the month, how many entries are touched by hand, and what audit trail remains when the machine proposes the entry. That last one is what separates a product from a demo, because during an inspection, the explanation of why that entry is there is still yours, not the provider's. Source

Teams with agents triple their output; what they deliver is another story

Linear published data from its platform: teams with connected code agents have gone from 21 to 65 weekly pull requests in two years, while those without them remain flat between 8 and 10; the total per workspace is up 111% over June 2024, and AI now writes just under half of everything created in the tool—about 2,500 issues per week in August compared to virtually none in mid-2024. For your company: the number that matters isn't the 111%, but the distance between 65 and 8, and the nuance that accompanies it: tripling what is produced does not triple what reaches production; it moves the bottleneck to review. Link this with today's first news story and you'll see the full problem: if half the work is written by a machine and the side looking for bugs is also automated, reviewing ceases to be a courtesy step. The practical consequence is budgetary: if you're going to put agents in your team, budget for review—people, time, and automated analysis—at the same time as the licenses, not six months later when no one knows who wrote what. Source

What to watch tomorrow?

Whether OpenAI resumes that training or not, because that schedule will determine the agent tools they'll be selling you next year. And whether Google and Microsoft mimic Anthropic by letting their assistants send email without confirmation: the day they do, having your company policy written ceases to be a best practice and becomes urgent.