noticias ia para empresas agentes de ia anthropic

What's holding AI back inside companies is no longer the model but the people who know how to get it into production: Anthropic is putting up $100 million to train 10,000 engineers who deploy it from the inside. Meanwhile, agents already create seven in ten new databases on Supabase, which raised $150 million and is buying Turso, and that's why permissions are becoming the critical piece: Apple will tighten Full Disk Access on the Mac. And an AI-found flaw was under attack the day after it went public. Picks up from the Saturday, October 3 radar.

Anthropic puts $100 million into training 10,000 engineers who take AI into production: what you're missing isn't the model, it's the person

Anthropic launched the Claude Frontier Academy on Friday, with a $100 million commitment to train 10,000 "frontier deployed engineers" by the end of 2027: the engineers who embed inside a company and turn an idea built on Claude into a system running in production. The first cohorts come from Accenture, Deloitte, McKinsey, Morgan Stanley, and Novo Nordisk; entry is by nomination and classes are held in San Francisco, New York, and London. The method borrows from medicine: a few in-person days built around a simulated enterprise deployment, "from the right use case through security review to handover," a graded practical, and then a 12-week residency leading a real project at their own company, named from day one. The first credentials will arrive in early 2027. For your company: you won't get into that academy, but the diagnosis is free. If the company that builds the model is spending $100 million on people, your bottleneck isn't picking between ChatGPT and Claude either. You need one person—not ten thousand—who owns each process you automate, with a named project and a deadline, not a mandate to "explore AI." And if you bring in outside help, get the three stages Anthropic tests for in writing: what use case was chosen and why, what security review was done, and what the handover looks like, including documentation, access, and who maintains it if the vendor disappears. The consulting firms in the first cohort will show up with a new badge; ask them to show you one of their own processes running anyway. Without that person, no process climbs from INTEGRATED to UNATTENDED: it stays a demo. Source

Supabase raises $150 million and buys Turso because agents now create 7 in 10 new databases: limit what yours can delete

Supabase, the Postgres-based database where many apps built with Lovable, Claude Code or Codex end up, announced a $150 million round on Friday led by GIC, Singapore's sovereign wealth fund, with CapitalG (from Alphabet), IronArc, and SquarePeg. It comes four months after its $500 million Series F at a $10.5 billion valuation, also led by GIC, and part of the money will provide liquidity for employees. At the same time, it's acquiring Turso, which rewrote SQLite so a single server can manage millions of small databases, one per agent; it didn't disclose the price and promises that "for existing users, nothing changes." The number behind the deal: Supabase adds more than 4 million databases per month, and nearly 70% of the new ones are created by agents or AI tools, not people (they were already the majority in June). At its event that day, it rolled out, on by default, a step in which its MCP server asks for confirmation before an agent creates a paid project or runs a command that could delete data, and warned that it "isn't a guarantee"; new tokens are now limited by default to the projects and permissions you choose. For your company: if your website, internal app, or data dashboard was built by an agent, find out where the data is and who has the key. Three checks: each agent gets its own token, scoped to its project with only the permissions it needs (the new default doesn't touch the tokens you already have); automated reviews run read-only; and keep a backup out of the agent's reach, because confirmation before deleting is a brake, not a safety net. And the money read: databases for agents are a category that's consolidating, and whoever hosts you today could have a new owner tomorrow, just like Turso's customers. An agent with permission to delete everything isn't working UNATTENDED: it's working without a net. Source

Apple will tighten "Full Disk Access" on the Mac because of AI agents: check today which apps have it at your company

Apple warned developers on Friday that it will add controls to macOS "Full Disk Access," a permission meant to let backup apps work that largely sidesteps the system's privacy controls. According to Apple, some developers use it in ways that expose everything on the computer—files, email, messages, and even browsing history—without the user fully understanding it, and going forward it can only be granted with "very explicit user action." Apple gives the reason itself: "As AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially." No date was given. This comes days after an Inc. columnist reported that Muse, Meta's agent, knew the content of his messages; Meta responded that this requires enabling both Full Disk Access and its Messages connector. For your company: don't wait for the update. On every work Mac, System Settings → Privacy & Security → Full Disk Access shows which apps have it; backups and antivirus should be there, not the AI assistant someone installed to summarize their email. With an MDM, you can deny it by profile for selected apps across all devices. That email and those messages contain personal customer data: GDPR requires appropriate security measures (Article 32), and an agent reading everything without anyone signing off on it doesn't qualify. Put it in your AI use policy too: which agents, with which permissions. An agent that sees everything isn't INTEGRATED: it's running loose. Source

A flaw found by Mythos, Anthropic's AI, was under attack the day after it went public: if you can't patch it in hours, don't leave it exposed

A researcher from Horizon3, an AI penetration testing company, on Wednesday published a critical flaw he found with Mythos, Anthropic's bug-hunting model, which only select partners can access. It's in Rejetto HFS, a free open-source file server, and lets anyone on the internet become an administrator and run code on the machine (CVE-2026-61500, 9.3 out of 10). He published it with a video walking through the exploit, and by Thursday night VulnCheck was already seeing attacks from an IP address in China against servers in the US and Japan. According to VulnCheck's log, Mythos and Anthropic's Glasswing program have now turned up 286 flaws with their own CVE, and until Thursday only one had been exploited in the wild: this is the second. It affects versions 3.0.0 to 3.2.0 and is fixed in 3.2.1. For your company: if anyone is sharing files with HFS from an office computer, update it today; if that's not possible, take it offline. In Saturday's radar we reported how an agent broke into DIVD through two unknown flaws; here, just one day passed between the flaw going public and the attack, because every published finding is also a recipe. Two things: sign up for INCIBE's security alerts for businesses (Spain's national cybersecurity institute) or your country's equivalent to hear about it the same day, and whatever you can't patch in hours—the NAS, the file server, the camera panel—take it off the internet and access via VPN. Source

What to watch tomorrow

When Apple's new control ships and whether it forces users to re-grant the permission to apps that already have it, and whether more attacks show up on the 286 Mythos flaws. And a thread we're following: David Robinson, who led the writing of OpenAI's safety reports before leaving the company (Saturday flash), wrote in The Atlantic that he left because, in his view, the company's culture "is broken"; we'll see whether OpenAI responds.