noticias radar ia agentes de ia ciberseguridad automatizacion

The Friday radar was about fragility; today's is about speed. Palo Alto has published the autopsy of an intrusion in which AI agents executed every step and dismantled the cloud, identities, and deployment pipelines of an entire company in less than ten hours, while Tenable and OpenAI set up a security review for agent components that anyone can download from a public registry. In product news, Microsoft drops the transcription hour to ten cents and xAI opens its persistent agents to companies with audit controls. And in money news, Crusoe triples its valuation to 30 billion and Andreessen pays 300 million to break free from single-chip manufacturer inference.

Ten hours to dismantle a company: the detection window for an attack is no longer two weeks

Palo Alto Networks' research unit published the autopsy this week of an intrusion where a human attacker delegated tactical execution to AI agents, updating it on Thursday to specify it wasn't ransomware but a complete environment takeover. The agents entered through an internet-accessible web service and chained together more than fifty techniques from the MITRE ATT&CK catalog—reconnaissance, credential theft, privilege escalation, exploitation of deployment pipelines, and hijacking the victim's own AI infrastructure—working in parallel across cloud, identity, CI/CD, and containers simultaneously. They took less than ten hours for what would have taken a human operator about two weeks. There was no zero-day or elite craftsmanship: what made the difference was the operational efficiency of a system that monitored, evaluated, acted, and replanned in real time. Upon finishing, the attacker left the victim an eighty-page technical report detailing the dozens of flaws exploited. For your company: if your incident plan is a sequence of steps with email approvals, it's late by design. Write the simultaneous containment list today—revoking credentials, killing sessions, freezing deployments, and isolating cloud accounts all at once—with a primary and backup name for each action, and ask whoever manages your systems to watch for the signs described in the report: bursts of API requests and rapid jumps from authorization errors to correct responses. Source

The components connecting your agents to your systems come from a public registry no one was checking

Tenable and OpenAI announced Wednesday, at the OpenAI cybersecurity summit, a security review process for agents, skills, MCP servers, and multi-agent playbooks published on CyberAgents Exchange, the open registry Tenable launched in August which already hosts over a hundred community-submitted components. The inspector combines automated evaluation using OpenAI's cybersecurity models, exposure analysis from Tenable's own platform, and human review by their researchers; it will be available this month. Translated: until now, over a hundred pieces that people plug into their environments were circulating without anyone looking at them. For your company: every MCP server connecting your assistant to your CRM, email, or accounting is third-party code running with your credentials, and it was almost certainly installed by someone on your team in five minutes without telling anyone. Build the inventory sheet today: which component, who installed it, what systems it accesses, with what permissions, and when it was last reviewed. And set the rule before it's needed: no new connector enters without an internal owner, least privilege permissions, and a review date. Source

Transcribing an hour of calls now costs ten cents, but the price expires on December 31

Microsoft introduced MAI-Transcribe-2 on Wednesday, its new speech-to-text model, in public preview within Microsoft Foundry: $0.10 per hour of audio, 72% less than the previous generation, with 60 languages, a 5.2% average word error rate on the FLEURS standard, speaker diarization, word-level timestamps, vocabulary biasing for your terminology, and tolerance for language switching within the same sentence. Microsoft claims it is ten times faster than OpenAI's equivalent model and seven times faster than ElevenLabs'. The fine print: this rate is promotional and valid until December 31, 2026; the January rate has not been published. For your company: a thousand hours of sales and customer service calls transcribed costs a hundred dollars, so there is no longer an economic excuse to go without a record of what your customers are saying: that is step one of the ladder, and it is the material used to automate responses later. Two warnings. It's the second price this week that expires exactly on December 31—the other was Google's fast model—so budget 2027 at the normal rate, not the promotion. And always save the original audio: if the provider raises the price, migration only hurts if you threw away the raw material. Source

xAI opens its permanent agents to companies: two weeks free and three things to demand before giving them the keys

xAI opened Grok Bot to enterprise customers on Wednesday. These are permanent agents, each with its own cloud computer, that enter the company's web applications just as a person would, learn a workflow after being shown once, pass context between each other, and work twenty-four hours a day in sales, recruitment, marketing, finance, or engineering. New in this release are the controls: isolated environments per user, zero default access, and layers of access, network, and audit control. Grok and Cursor Enterprise customers get two weeks free and can invite the entire staff, including those without licenses. For your company: an agent working non-stop with your credentials is, for all practical purposes, an employee without a contract or a boss. If you enter the free trial, three non-negotiable conditions: a separate account for the agent and never a person's, read-only permissions for the first week, and a written list of actions that require human confirmation—payments, bank account changes, customer communications, and deletions. And ask in writing two things the announcement doesn't clarify: what is the data retention policy and is the audit log already functional or merely announced. Source

3 billion for Crusoe and 300 million for whoever frees inference from a single chip manufacturer

Crusoe, the data center operator serving OpenAI, Microsoft, and Meta, closed a round of over $3 billion on Thursday at a valuation near $30 billion, co-led by Atreides Management and Valor Equity Partners with Mubadala Capital participating. Ten months ago, it raised 1.38 billion at a 10 billion valuation: it has tripled in less than a year after signing a 13 billion, five-year cloud contract with trading firm Jane Street. The following day, Gimlet Labs announced a 300 million Series B at a 3 billion valuation, led by Andreessen Horowitz with Arm and Microsoft's investment arm joining the capital, six months after an 80 million round. Their product is multi-silicon inference: software that distributes model workloads across accelerators from different manufacturers instead of tying them to a single chipmaker. For your company: the big money this week didn't go to the application you use, but to the plumbing underneath, and one of the highest-paid bets was precisely to stop depending on a single hardware provider. The lesson for your contract is the same on a smaller scale: ask your AI provider which cloud and infrastructure they run on, if they can change it without moving your price, and what happens to your rate when the current commitment expires. The portability they are buying with 300 million is what you should have in writing in a clause. Source

What to watch for tomorrow?

If any other incident response team publishes its own autopsy of an intrusion executed by agents: with two documented cases, it stops being an anecdote and becomes the new reaction time you have to assume. And if the agent component inspector really comes out this month and what criteria it uses to reject a piece, which is what will decide if it is useful or just a rubber stamp.

Watch the 1-minute video