noticias ia para empresas agentes de ia ciberseguridad

Three days of news, one common thread: what's holding AI agents back is no longer the model, it's everything around it. At Anthropic, which has rebuilt the way it works around its own agents, those agents now log six hours for every hour its employees work, and the bottleneck is human attention. Proofpoint is paying $6.3 billion for Varonis to control what data each person and each AI can see. Banks, airlines, and stores are closing their doors to agents browsing their websites, and OpenCode warns that the OpenAI and Anthropic API keys stored by its users may have been exposed. Previous edition: Friday's radar (Oct 16).

At Anthropic, AI agents already work six hours for every human hour: the bottleneck is the reviewer, not the machine

Anthropic has published a study on how work has changed within the company, drawing on internal data on headcount, AI use, output, calendars and hiring, plus 17 interviews. By August, the agents its employees launch were working six times as many hours as their human handlers. Code output per engineer is running at five times its pace of six months ago, and outside of technical roles, lawyers review more contracts, accounts payable processes more invoices, and sales reps close more deals; a financial manager, for example, built an agent to review purchase requests. Junior roles have fallen from 27% to 9% of its technical hires. But the most useful number is a different one: in interactive sessions, agents spend about 80% of their time waiting for a person to reply, and reading everything they produce would take up more than half a work week even for non-technical teams. The response: written standing instructions for agents have grown from a few hundred to 8,500 in a year. These are internal numbers from the company that sells the agents, and Anthropic is not a small business. For your business: the next step up isn't buying more agents, it's to stop being the bottleneck. List the tasks where the agent waits for your sign-off and decide which ones it can close on its own within a clear limit (an amount, a customer type, something reversible); write down once how you want things done, instead of repeating it in every conversation, and ask it to show you only the exceptions. If every result waits for you to read it in full, you are still ASSISTED even if you have ten agents; only when you decide what you don't need to look at can a process move to UNATTENDED. Source

Proofpoint pays $6.3 billion for Varonis: controlling what data your company's AI can read is now a consolidating business

Proofpoint, owned by private equity firm Thoma Bravo, has agreed to buy Varonis for approximately $6.3 billion in cash: $49.75 per share, a 52% premium to its June share price, before the first sale rumors, and about 8.4 times its trailing twelve-month revenue. Varonis sells visibility into data: where sensitive information is, who can open it, and how it's used. Thoma Bravo is providing the equity and has lined up the debt financing; the deal is expected to close in the first half of 2027 if approved by shareholders and regulators. The resulting company, with over 16,000 corporate clients, promises to control what people and AI can access. Proofpoint CEO Sumit Dhawan put it this way: "As AI becomes embedded across the enterprise, the data it can access determines both its value and its risk." For your business: the money is going into controlling what AI can see, because of the problem we covered on Wednesday the 14th: an assistant connected to your email or folders can read everything the employee using it can read. Before connecting Copilot, Gemini, or Claude to your documents, run three searches: files shared with "anyone with the link," folders open to the entire organization, and permissions still active for former employees or external collaborators. If you work with Varonis, your provider is changing owners: watch your next renewal closely. Only once permissions have been reviewed can the assistant move up to INTEGRATED; until then, connecting it to your folders is handing out keys you don't know exist. Source

Banks, airlines, and stores shut the door on AI agents: don't build a process that depends on someone else's website

Users of Meta's Muse, xAI's Grok, OpenAI's Dots, and Instinct are hitting blocks on major websites. Yelp only lets in agents that pay for its data licensing program, Delta prevents them from booking tickets, Amazon says Muse's continued access violates its site policies, and JPMorgan Chase and Charles Schwab are blocking agents trying to access customer accounts. Schwab has spelled out its position: it bans agents that use a client's login credentials or automate the browser, lets certain language models connect to Schwab to trade, but not to move money, and pushes its own assistant, Charley. One consultant summed it up: "An outside agent works for the client. Charley works for Schwab." Meta counters that "turning away a personal agent means turning away the customer behind it" and has signed deals with Walmart, Gap, Sephora, and Best Buy so Muse can shop on their sites. Meanwhile, anti-bot systems are stopping many agents even if they haven't been specifically banned. For your business, two takeaways. If you automate purchases, bookings, or supplier queries with an agent browsing external websites, it could stop working any day without notice: for anything that has to run unattended, ask the provider for an API, an official connector, or an MCP server, and leave browser agents for one-off tasks. And if you are the one selling, revisit the list from Thursday the 8th: what an agent can query without identifying itself, what requires the customer to log in, and what you will never do without the person. A process that breaks when another company updates its anti-bot isn't UNATTENDED: it's on borrowed time. Source

Unauthorized access to OpenCode leaves OpenAI and Anthropic API keys exposed: one key per tool, with spending limits and alerts

OpenCode, an open-source coding agent where each developer connects their own OpenAI, Anthropic, or Google keys, has warned users that someone gained access to a legacy system it was decommissioning and that keys stored in their workspaces may have been exposed. It affects 1.4% of its Console users, those with external providers linked, and the company ties it to a large-scale distillation attack: gathering responses from one model to train another. It's asking users to revoke and replace their keys and review recent usage. Some users are already reporting unauthorized charges, such as $400 on an Anthropic account or about €200 on a Google account, and one received usage alerts from OpenAI and Anthropic before the OpenCode notice arrived. For your business: the risk extends to any tool where you have pasted an API key, whether it's a coding agent, an automation tool like n8n or Make, or a spreadsheet add-on. Today, take stock of where every key lives, use one per tool and per project so you can revoke one without stopping everything, and set a spending limit with alerts for each, like the caps we proposed on Thursday the 15th. Running all your automation on a single key isn't INTEGRATED: it's one key for the whole house. Source

What to watch this week

Whether OpenCode publishes technical details of the attack and how many keys were actually used; whether more banks follow Schwab and open a controlled path for agents while closing browser access; and on Friday the 23rd, the 4GB version of Nemotron 3.5 Bolt that Nvidia announced for that day, which we covered on Friday the 16th.