AI Radar for Companies — Friday, November 20, 2026
· CompaniesAutomation
Anthropic nears $2 trillion on the Nasdaq, Palo Alto pays $4.5B for Applied Compute, AI agents hit the European Commission and Google ships Gemini 4 Neon.
Anthropic is now a public company and touched a valuation of about $2 trillion at the open, while Palo Alto Networks is paying $4.5 billion for a startup that tunes open models for businesses: the money is backing the big vendor and the alternative at the same time. The other story of the day is security. The European Commission has spent more than ten days rebuilding systems after an attack carried out with an AI agent toolkit, and Microsoft's official Python package for building agents shipped password-stealing malware. Google launched Gemini 4 Neon, cheaper than its top models, and in Spain 61% of companies now use AI but only 36% have trained their staff. This follows Thursday's radar, the one on AI coding at 5% of the price.
Anthropic opens 34% above its IPO price and nears $2 trillion: your AI vendor now has to please shareholders every quarter
Anthropic shares, priced at $83, opened on the Nasdaq on Thursday at $111.50, up 34%, valuing the company at around $2 trillion. The 980 million-share listing is the biggest ever, according to the US press. In an interview with CNBC, CEO Dario Amodei admitted that the company's compute needs are hard to forecast and that it tends to err on the low side, despite the $900 billion in commitments it disclosed to the SEC. He also said Anthropic can compete "both at and away from the frontier" with cheaper models. On the same day, Mozilla pointed to its own research this year: open models already handle a third of all tokens, and the gap with the best closed models is down to 3.3%. For your business: your vendor going public doesn't change your contract today, but it does change its incentives. As we said in Thursday's flash, it now reports every three months. Your best protection is having at least one workflow that runs just as well on an alternative model, even an open one. That's the INTEGRATED rung of the ladder, and it's what gives you leverage when the price changes come. Source
Hackers hit the European Commission with an AI agent toolkit, leaving staff unable to manage payroll or time off for more than ten days
The European Commission has been dealing with the fallout of an attack on its internal systems for more than ten days. Its security team found the attackers were using ARTEX, a publicly available penetration testing framework that launches AI agents able to plan, probe for weaknesses and move through a network with almost no human input. The Commission had to take down Sysper, the app staff use for payslips, timesheets and vacation requests, and MyPMO, which handles medical reimbursements. Both came back this week, but Sysper went down again on Wednesday. The Commission hasn't said whether personal data was stolen. The ARTEX developer pulled the code in October, but copies were already circulating, and last week the FBI counted more than 1,600 ARTEX consoles exposed online, with cases where attackers went from picking a target to privileged access in under three hours. For your business: an attack that used to take weeks now takes hours, and it doesn't need an expert. Check today which of your services are exposed to the internet, turn on two-factor authentication for every admin account, and decide how you'd run payroll if your HR app went down for a week. We saw in the November 8 radar that ARTEX was already slipping its limits. Source
Microsoft's official Python package for building agents installed a password stealer: if you use it, rotate your keys
Versions 1.26.1 and 1.26.2 of agent-framework, Microsoft's Python library for building AI agents, with nearly 200,000 downloads a month, were published with malicious code through Microsoft's own official account on the PyPI repository. You don't even have to use the library to get infected: having it installed and starting Python is enough. The code steals environment variables and credential files for Azure, Git, Docker and Hugging Face, and it also requests fresh tokens from GitHub Actions, Azure DevOps and Azure and sends them out immediately. Microsoft has made several repositories private while it investigates. For your business: if your team or your vendor builds agents with this library, ask today which version is installed on laptops and build servers. If either of those two shows up, assume every key on that machine is stolen and replace it. And going forward: pin versions in every project, no automatic updates on servers with access to production, and keys that expire on their own. Source
Palo Alto Networks pays $4.5 billion for Applied Compute: your cybersecurity vendor wants to train your AI models too
Palo Alto Networks, one of the world's largest firewall makers, will buy Applied Compute for $4.5 billion in stock. The startup tunes and serves open-weight AI models for businesses, in the public cloud or in their own environments, and the deal is expected to close between February and April 2027. CEO Nikesh Arora said open models have gone from under 5% to more than 60% of token traffic in a year, without saying where the figure comes from, and that Palo Alto wants more than 100 megawatts of its own compute within three to five years. It comes after the $25 billion purchase of CyberArk, the identity security specialist. As a Greyhound Research analyst put it, the guard wants to build the vault it guards. For your business: the market for open models tuned to your company is consolidating inside the big security vendors. If your security partner offers to train a model for you, get it in writing before you sign that the trained weights are yours and exportable, that they won't train on your data without consent, and that you can walk away without penalty if the company changes hands. Source
Google launches Gemini 4 Neon, cheaper than its top models and first on an automation benchmark: pick the model yourself before Google picks it for you
Google unveiled Gemini 4 Neon on Wednesday night, built for agents that work for hours across many tools. It costs $3 per million input tokens and $15 per million output tokens, cached text is 95% cheaper, and it handles a one million-token context. Google says it ranks first on AutomationBench, Zapier's automation test, at 56.3%, and leads several finance, accounting and spreadsheet benchmarks. It's already in the Gemini API, Gemini for business and Antigravity, Google's coding environment, where an orchestrator swaps models on the fly depending on the task. Its top models, Carbon and Argon, will cost $4 per million input tokens and $20 per million output once the discount ends next month. For your business: a mid-tier model that scores like an expensive one cuts the cost of automating reports and reconciliations. But if you let the orchestrator choose the model, it also chooses your bill. Try Neon on one repetitive task, pin the model in your settings and compare the cost per finished task with what you use today. Source · Source
61% of Spanish companies now use AI, but only 36% have trained their staff and 45% have no usage policy at all
InfoJobs' annual survey on digital transformation, published on Thursday, found that 61% of Spanish companies say they use some kind of AI tool, up from 38% in 2025 and 24% in 2024; counting those that use it without realizing, the figure rises to 77%. Only 36% trained their employees in the past year: 64% of large companies versus 22% of micro businesses. 45% have no defined policy on how AI is used at work, and among those that set limits, the top reason is data privacy and security, now a concern for 64%. 28% have built their own AI tools. For your business: if your team uses ChatGPT or Copilot with no rules, you're on the ASSISTED rung of the ladder, with no safety net. The cheapest step up to INTEGRATED is a one-page policy covering which data never goes into a chat and which tools are approved, plus two hours of training on real tasks from your business. Source
What to watch tomorrow
The close of Ignite this Friday, November 20, with the price of Copilot Autopilot still pending, and Microsoft's explanation of how its official PyPI account was compromised; plus whether the European Commission confirms or rules out the theft of staff data.