noticias ia para empresas ciberseguridad agentes de ia

The week closes with two warnings about the AI tools teams already rely on. A tampered build of the Claude Code extension sat in Cursor's extension store for about eleven hours and may have reached up to 5,000 users, and Chainalysis, the firm that tracks crypto thieves, says attackers using AI agents got into its systems and took customers' names, emails and job roles. The money sends its own signal: S&P moved Meta's outlook to negative because it's spending on AI with no clear date for getting paid back, while Baseten, which sells the compute to run models, is set to more than double its valuation. And there's an awkward question for anyone buying agents: how many of the tasks end up with a person? Catch up on Saturday's radar.

A fake Claude Code slipped into Cursor's store: every extension you install is also a door

On Thursday, October 29, for roughly eleven hours, the extension marketplace of Cursor, the AI code editor, served two tampered versions of the Claude Code extension (2.1.312 and 2.1.313). Cursor says up to 5,000 users may have been affected. According to Microsoft's analysis, the injected code ran as soon as the extension loaded: it collected the machine name, the user name, the Git email, and the names (not the values) of up to forty environment variables containing words like KEY, TOKEN or PASS. It then added the attacker's SSH key, tried to open a reverse tunnel for remote access, and dropped startup scripts named to look like system components. Cursor says bugs in the malware and missing admin rights blocked remote access on most Windows and Mac machines, but the data collection and the persistence could still have landed, and uninstalling the extension isn't enough. Anthropic wasn't breached, and neither the Visual Studio Marketplace nor the official package was affected; only Cursor's store was, and Cursor has pulled the versions, rotated its keys and rebuilt the service. The takeaway: extensions update themselves and nobody reviews them, which makes them the shortest path to the laptop of whoever holds the keys. Find out today who uses Cursor and whether they updated Claude Code on Thursday; on those machines, look for the "svc-telemetry" key and rotate credentials. Then set one rule for every AI tool you've plugged in, which is what the INTEGRATED rung on the ladder is about: a closed list of approved extensions, installed from the official source, and no secrets sitting in a laptop's environment variables. Source · Source

S&P warns Meta over AI spending it can't yet bill for, while Baseten, which bills for running models, is set to double its value

On Friday, S&P Global cut Meta's outlook from stable to negative while keeping its AA- rating, citing "uncertainty" over its capital spending and cash flow. Meta expects to spend $135 billion to $145 billion this year, posted negative free cash flow of $2.12 billion in the third quarter, its first quarter in the red, and plans to raise $95 billion in equity to fund the build-out. The agency questioned "Meta's ability to monetize its AI investments" relative to its peers and said it's unclear when its Muse agent will bring in more than it costs; Meta said it "respectfully disagree[s]". The same evening, Bloomberg reported that Baseten, which rents out the infrastructure to run AI models and charges for that usage, is set to raise a round in two tranches that would value it at $26 billion and $30 billion, more than double its June mark. It's the same week OpenAI scaled back its own round, as we covered in Saturday's flash edition. The takeaway: investors are now separating the companies that spend on AI from the ones that get paid for it, and that pushes every vendor toward usage-based billing. If you rely on a tech giant's free models or tools, assume someone is working out how to charge for them. Work out today what each automated task costs you, in euros or dollars per invoice or per query, and pick the model per task. If you don't measure it, you can't defend it when the price changes. Source · Source

Some AI agents are quietly backed by people: ask how much a human does before you sign

CNBC reports that several personal-agent startups hand off to people the jobs AI still fumbles, especially phone calls, appointments and CAPTCHAs. Dazzle AI, run by Marissa Mayer, routes some requests to human contractors: its agent once mixed up window glass with a glass tabletop, and a person sorted it out in seconds, where the agent "would have burned a ridiculous number of tokens". Duckbill has launched an API that lets other agent companies send the tasks their agents are likely to fail to human workers, broken into small pieces, and Wajo AI uses people for calls and CAPTCHAs. Meta previously tested contractors who made calls on behalf of its Muse agent, though it says it doesn't use them today. The takeaway: a human backstop isn't cheating, it's design, but you need to know about it. If you buy an agent service, get it in writing: what share of tasks ends up with a person, who those people are and where they work. Under the GDPR, those people are handling your data, and the vendor can't subcontract without your authorization. And in your own processes, being UNATTENDED doesn't mean having no people involved; it means knowing which task goes to whom, and measuring it. Source

Chainalysis hit by AI agents: what the attackers take is your staff directory, and phishing follows

Chainalysis said on Friday that this month it detected an outside attacker using "offensive agentic techniques" to break into its systems and internal tools. According to the email it sent customers, it has found no sign of access to investigation content, plain-text passwords, MFA secrets or session tokens, but user IDs, full names, email addresses and job-role labels for people set up on its platform were exposed. It's telling users to watch for password-reset requests they didn't make and to turn on two-factor authentication, and it says the incident has nothing to do with the Binance theft. The takeaway: names, emails and job titles are exactly what a scammer needs to write a convincing email to whoever approves your payments. If a vendor tells you it's been breached, warn the people on your team who appear in its records that same day, and require two-factor authentication on every cloud tool you use. Your vendor has to tell you without undue delay, and in the EU you then have 72 hours to report it to your data protection authority if people's data is at risk. Source

What to watch tomorrow

Whether Cursor confirms how many machines actually ran the malicious code, and the customer-by-customer breakdown Chainalysis has promised. Plus whether Meta answers S&P with spending cuts or new paid products.