noticias radar ia agentes de ia seguridad ia automatizacion

Yesterday's radar was about who holds the keys; today it's about who pushes the button. Starting today, Claude Code launches in automatic mode for Pro, Max, and Team plans: it stops asking for permission command by command and filters each action with a classifier that, in Anthropic's tests, stopped 89% of dangerous orders, compared to 13.6% stopped by humans. In parallel, SpaceXAI released Grok 4.6 at the same price as the previous version with a focus on long-running agents, and a fund with OpenAI backing raised $2 billion to buy accounting firms and IT companies to rewrite them with AI.

Starting today, coding AI stops asking you for permission step-by-step — and the data says it's the right move

Anthropic announced on August 7 that as of today, August 14, new Claude Code sessions in Pro, Max, and Team plans start in automatic mode: instead of interrupting you to approve every command, the agent passes each action through a classifier trained to block anything irreversible, destructive, or targeting outside your environment; when it blocks something, it looks for a safer path or asks you. The reason lies in a study with 1,053 paid testers: humans only detected 13.6% of dangerous commands while the classifier caught 89%, and human attention degraded as the session progressed — from 17% at the beginning to less than 5% after fifty messages. Anthropic is also stopping charges for extra classifier tokens, and automatic mode remains optional in Claude Enterprise, the API, Bedrock, Google Cloud, and Microsoft Foundry. For your company: the data that matters isn't just that the machine reviews better, it's why — approving two hundred times in a row turns the click into a reflex — so stopping the reliance on human vigilance is sensible, but only if the environment is sandboxed. If someone on your team uses this tool, the default behavior changes today: check which machine it runs on, with what credentials, and if production or customer data can be reached from there. And if you have an Enterprise or API plan, the change hasn't reached you yet: take the opportunity to set your default mode in writing before it does. Source

Yesterday's bill now buys an agent that handles longer tasks

SpaceXAI released Grok 4.6 on August 12, a retraining of version 4.5 focused on long-running agents and visual work, and kept it at the same price: $2 per million input tokens and $6 per million output tokens, double for the fast variant and starting at 200,000 context tokens. It ties with GPT-5.6 Sol on the Artificial Analysis index (61 points), leads CursorBench 3.2 with 69.9% and DeepSWE with 65.9%, but stays at 26% on Terminal-Bench; it is available in Cursor, Grok Build, and via API. For your company: the takeaway is about budget, not rankings. The price of top-tier tokens has been flat for months while capacity for chained tasks rises, so if you have an automation built on a six-month-old model, you are paying the same for less. The test you need to do isn't reading benchmarks: take ten real cases from your process, run them through the new model, and compare the result and cost. And watch for the asymmetry — the same model that leads in generating a first prototype fails in terminal tasks: measure against your specific task, never on the average. Source

Big money is already buying accounting and IT firms to rewrite them from the inside with AI

Thrive Holdings — a spin-off of Thrive Capital, with OpenAI as a shareholder since December 2025 — closed a $2 billion round on August 12 at a $12 billion valuation, with SoftBank, D1 Capital, and Altimeter. Their model isn't selling software: they buy traditional service companies and embed AI inside them. Their accounting platform, Current, already groups more than 50 firms and 2,000 professionals, and their tax product processed more than 7,000 returns with 98% accuracy and 30% less time; their IT vertical, Shield, brings together about twenty companies and claims to have multiplied the resolution speed of support services by 36. That's more than 70 companies in total, and part of the new money goes to a third vertical for infrastructure regulatory filings. For your company: this is no longer a tool you can buy; it's a competitor with a different cost structure. If you sell professional services by the hour — consulting, accounting, IT support, filings — whoever buys your neighbor and installs this inside will be able to lower prices without touching margins. Two tasks for this week: separate your billable hours between repeatable processes and professional judgment, and start automating the former yourself, because the competitive edge lies in doing it before whoever comes to buy you. Source

While you decide what permissions to give an agent, they keep getting in via a call to the IT guy

Uber Freight confirmed on August 12 that it is investigating a security incident after the Helix group — one of the aliases Google tracks under actor UNC6671 — published that it had exfiltrated email inboxes, cloud storage units, accounts payable files, and shipping documentation, close to one million files; TechCrunch reviewed emails with customers dated mid-June, and the company maintains that operations have not been affected. The group's entry method, described as rudimentary and very effective, is calling the IT support service pretending to be an employee and asking for a password reset; with this technique, the collective has collected at least $10.6 million in ransoms between January and May. For your company: you can have the best permission control for your agents and lose everything because someone polite called asking for a new password. Write your credential reset rule today: never from an incoming call, always with verification through a second channel that you initiate, and with a log of who requested it and who approved it. With AI-cloned voices, "I recognized their voice" is no longer a valid check. Source

What to watch for tomorrow?

Anthropic is negotiating to buy Decart for about $6 billion, a company whose software makes training and operating models cheaper by squeezing more out of chips: if confirmed, what moves in the medium term is everyone's token bill. And watch the calendar for automatic mode in Claude Enterprise and the API, expected next month: if you manage those accounts, better to decide your policy before the change decides it for you.

Watch the 1-minute video