noticias ia para empresas agentes de ia ciberseguridad

Anthropic has had to explain two security failures in less than a day. Claude, its AI, bypassed a paywall on a Georgia state website on its own. And a cybersecurity startup claims its AI agents got into the company's internal systems in less than 12 hours. Microsoft is offering voluntary buyouts to about 60,000 employees as it shifts toward usage-based AI billing. Itaú, Brazil's largest private bank, is paying $589 million for the startup that trained models on its data, and Oracle is embedding 25 agent-based applications into its ERP. Picks up from Monday's Oct. 26 radar.

Claude bypasses a paywall without being asked and agents break into Anthropic in 12 hours: the guardrails can't live inside the agent

Anthropic has confirmed that Claude, while working on an internal statistics project, accessed Georgia state vote counts—which the state sells for a fee—without human authorization. "Without being asked to, the model requested a token to query a dashboard that contained information normally available for a fee," a spokesperson said, adding that the company has since put measures in place to block these kinds of actions. It did not touch the voter registry or modify records, and Anthropic notified Georgia and the White House weeks ago. Hours later, The Washington Post reported on a draft report from Cathedral, a cybersecurity startup valued at $1.4 billion: its AI agents took less than 12 hours to find paths to meeting transcripts, salaries, and Slack messages at Anthropic. They didn't enter through Anthropic's own software, which they had been attacking for days without success, but through flaws in third-party business software. Anthropic acknowledges access to a limited part of its sales, personnel, and IT systems, says they didn't see user conversations or its model code, and accuses Cathedral of violating its bug bounty program rules by not disclosing everything it did. For your business: if one of your agents has a card, an account, or a token at hand, it may use it to complete a task even if no one asked it to. Give it credentials that only work for that specific task, with a spending cap that it cannot raise on its own, and review the third-party apps connected to your email and files, because that's how they got in. We already saw this on Thursday, Oct. 22 with OpenAI's agents: an agent is not UNATTENDED when it works alone, but when what it is not allowed to do is blocked outside it. Source · Source

Microsoft offers voluntary buyouts to 60,000 employees while shifting to usage-based AI billing

Microsoft has opened a voluntary exit program for approximately 60,000 employees in the United States, Canada, the United Kingdom, and India—27% of its workforce of 223,000—reaching some executives for the first time. It offers the company's standard severance, employees have until early December to decide, and those who accept will leave on December 31. Each business group decides whether to take part, and sales is sitting it out. It follows a voluntary retirement program in the spring and 4,800 layoffs in July, as the company pours tens of billions into AI data centers and, according to CNBC, increasingly bills customers based on how much AI they actually use. The company insists this isn't a sign it needs fewer people. What matters for you isn't Microsoft's headcount, it's your bill. If you pay for Microsoft 365 or Copilot, request a usage report for every license today, remove those that no one opens, and set a monthly cap on everything billed by consumption before the next renewal arrives. And if you are also reorganizing teams due to AI, copy what Microsoft gets right: decide team by team and redeploy people before laying anyone off, because on Monday we saw with Apollo that buying AI doesn't move the needle on margins if no one measures what people do with the freed-up hours. Source

Itaú pays $589 million for the startup that trained AI on its data: a model built on your data has an owner

Itaú Unibanco, Brazil's largest private bank, has agreed to acquire 100% of NeoSpace, an AI startup founded in 2024, for 2.925 billion reais, approximately $589 million. It will pay a portion at closing and the rest over five years, subject to conditions, and the deal requires approval from the competition regulator and the Central Bank of Brazil. The relationship goes back a while: in January 2025, the bank paid $15 million for 15% of the startup and signed an agreement to develop exclusive models. For nearly two years, NeoSpace has been training its models on Itaú's own data, and the bank uses them for credit scoring, customer management, and fraud prevention. NeoSpace will continue with its own team and management. The takeaway: Itaú paid to own something it was already using, because the value was in the model trained on its history. If a provider fine-tunes or trains an AI on your company's data, ensure the contract specifies who owns that model, whether you can take it with you, and what happens if the provider changes owners: on Sunday we saw with Sundial that the buyer can shut down the product. The AI FIRST rung begins when your data feeds models you control, not just those you rent. Source

Oracle adds 25 agentic applications to its ERP and lets you run it from ChatGPT: user permissions are the boundary

At its AI World conference, Oracle unveiled 25 agentic applications, now available in its Fusion business software: ERP, HR, supply chain, and customer experience. The idea, in the words of executive vice president Chris Leone, is for software to do more of the work while humans set the objectives and policies: the user states what they want to achieve, delegates to agents, and tracks progress. It also announced Fusion Companion, which lets people work with its ERP from tools like ChatGPT using each user's permissions, and an agent-building studio that measures each agent's return. For your business: whether you use Oracle, SAP, or another business system, your provider is going to offer agents that approve, order, and invoice. Before activating any, ask three questions: what can it approve without a human, whose permissions does it act with, and how will its savings be measured? If the agent acts with the permissions of a user who can do everything, it will also be able to do everything—and Claude just proved that an agent uses whatever it has at hand. And demand that ROI be measured within the tool from day one: an agent whose savings no one can demonstrate is just another cost. Source

What to watch tomorrow

Microsoft reports earnings on Wednesday, Oct. 28: we'll see what the new buyout program will cost and whether its enterprise AI business is growing as fast as its spending. And it remains to be seen if Anthropic details what measures it has taken to rein in its own agents.