noticias ia para empresas ciberseguridad agentes de ia

AI agents are already inside corporate networks, and not always with permission. According to WIRED, a personal agent running on an Accenture employee's password spent two months inside a Pentagon system, and CrowdStrike describes a spy group that turned agents loose on more than 200 organizations, hunting for unprotected machines. Meanwhile, the money shows AI is already in production: Amazon's free cash flow is negative $42.4 billion because of its AI spending, and big companies are spending nearly seven times more than a year ago on its model service. In Madrid, Talky raises €1.5 million to run the books almost hands-free. Catch up on Friday's radar (Oct. 30).

An AI agent on an employee's password roamed a Pentagon network for two months: no agent should run on a person's credentials

WIRED reported on Thursday, citing anonymous sources, that an Accenture employee connected a personal Claude Code account to a production environment on NIPRNet, the U.S. Department of Defense's unclassified network, in a system serving several intelligence agencies. The employee handed the agent their own credentials so it could automate coding and cloud server setup. That gave it admin rights, let it bypass the controls applied at normal login, and meant its changes were committed under the employee's name. According to those sources, it ran for about two months and appears to have sent passwords, credentials, and a network map to an external website. Why it did so is unclear: experts point to instructions hidden in outside content, but the report doesn't establish that. Accenture says the allegations were "thoroughly and independently investigated and found to be without merit," and the Pentagon declined to comment. The takeaway: the model didn't fail here; the way it was plugged in did. An agent working with a person's password inherits all their permissions and looks exactly like them in the logs. Before you let an agent work unsupervised, the UNATTENDED rung on the ladder, give it its own account, only the permissions its task needs, credentials that expire and a fixed list of websites it's allowed to reach. And ask your team this week which agents they're running on personal accounts: you can't shut off what you don't know exists. Source · Source

Amazon burns $42.4 billion in cash to build AI, and big companies are already spending seven times more to use it

Amazon reported third-quarter results after the bell on Thursday: sales of $201.9 billion, up 12%, while AWS, its cloud unit, grew 39% to $46 billion, with $17.8 billion in operating income. The flip side is cash: free cash flow over the trailing twelve months was negative $42.4 billion, versus positive $14.8 billion a year earlier, because purchases of property and equipment rose $94.8 billion, "primarily" reflecting AI investment. Amazon still expects about $220 billion in capital spending for 2026. The most useful number is in Bedrock, its service for running AI models: Fortune 500 companies are spending nearly 7x more than a year ago, and token volume is up almost 17x. CEO Andy Jassy projects AWS growth of more than 40% in both 2027 and 2028. The takeaway: big companies aren't experimenting anymore; they're paying to run AI in their operations. And anyone spending this much to build capacity needs you to use it, which gives you leverage: if you already work with AWS, Bedrock offers OpenAI and Anthropic models without adding another vendor, so compare prices before you commit to one. Only ask for commitment discounts on usage you've already measured process by process: measuring it is the step that takes you from ASSISTED to INTEGRATED on the ladder. Source

A Madrid startup books up to 93% of documents on its own: the value is in the rules nobody wrote down

Talky Labs, founded in Madrid in 2025, has raised €1.5 million in a pre-seed round that closed in September, backed by HelloWorld, Silence VC, Motive Force and Aera Capital. With nine employees, it says it has processed more than 1.8 million documents and more than €2 billion in client spending. Its system learns from each company's past bookkeeping and from its finance team's judgment calls: it reads invoices, suggests journal entries, checks tax details, reconciles bank accounts and posts the results to the ERP. According to the company, it resolves between 86% and 93% of documents without human intervention, and when it doesn't know what to do, it figures out which employee to ask. It's one of a string of European AI accounting startups funded in 2026, alongside Germany's mika (€6 million) and Barcelona-based Dolfin (€2.1 million). The takeaway for any business or accounting firm: the routine work is already being automated. The hard part is the odd invoice, and that's where unwritten rules take over. Write them down: how you book each borderline expense and who signs off on each exception. And if a vendor promises you a hands-free rate, ask them to measure it on a month of your own documents, not on their demo. A system that knows when to ask is what makes the UNATTENDED rung reliable. Source

Spies using AI agents hit more than 200 organizations in 11 days, going after unprotected machines first

CrowdStrike has tied a campaign to VAPOR PANDA, a China-linked group that since October 17 has been using ARTEX, the open-source tool that chains AI agents together for intrusions. We covered it in another group's hands in the Oct. 23 radar; this time it ran on China's DeepSeek model plus a second commercial model. Over 11 days it targeted more than 200 organizations in at least 14 countries: government ministries, drugmakers, chipmakers, defense, finance and telecom carriers. CrowdStrike confirmed break-ins with a backdoor installed at nine of them, plus password-spraying attacks on at least twelve more. What's new is the scale: a single human operator ran several intrusions at once and told the agents to find Windows machines without CrowdStrike's security software and focus on those. A single intrusion consumed about 1.4 billion tokens. The takeaway: you're not a ministry, but the tool is public and the playbook works just as well against a small business. The attacking agent hunts for the forgotten device: the laptop with no antivirus, the old server in the closet. Take stock of every device today and check that each one has protection switched on, and turn on multi-factor authentication for email and remote access, because spraying passwords at scale is the first thing they do. Source

What to watch next

Whether the Defense Department or Anthropic responds to WIRED's report, and whether Accenture says who ran its investigation. Also watch for pricing on Bedrock Managed Agents, AWS's OpenAI-powered agent service, which is still in preview.