noticias ia para empresas mistral agentes de ia

Europe now has an AI model in the big leagues that, once its weights are published, no one will be able to withdraw: Mistral launches Large 4, with 1 trillion parameters. SAP buys TechWolf, the AI that maps what every employee does and knows, and cybersecurity money is flowing to agents that attack before criminals do: Hadrian raises $40 million. And insurers are starting to wonder who pays when an agent goes rogue. Picks up from Tuesday's radar, October 6.

Mistral launches Large 4, a 1-trillion-parameter European AI with open weights: a plan B that no one can take away from you

France's Mistral released a public preview of Mistral Large 4 on Tuesday, its largest model yet: 1 trillion parameters, of which it uses 49 billion for each task, and it understands both text and images. It is already available via API at $1.36 per million input tokens and $4.18 per million output tokens, and the weights—the full model to install wherever you want—will arrive at the end of October. Mistral trained it from scratch on 3,800 Nvidia chips in its own European data centers, and offers a European deployment it runs end to end, "under European law." It's a leap, but not the top spot: in the independent Artificial Analysis index, it jumps from its predecessor's 9 points to 38, still far from Claude Opus 5.5's 58. Its big selling point is cybersecurity: on a test that asks a model to reproduce and patch a real vulnerability, it scores 82%, while Claude Opus 5.5 and GPT-6 Astra stay near zero because they refuse to perform the task. You "cannot afford to be vulnerable to the fact that the model you are using to protect yourself might disappear one day," Guillaume Lample, co-founder of Mistral, told The Deep View. For your company: open weights don't mean you'll install a 1-trillion-parameter model in the office—it needs a data center—but rather that your integrator or a European cloud will be able to serve you the exact same model without anyone changing or withdrawing it. Three things. Per token, it's cheaper than Claude Sonnet 5.5 ($2 and $10), but, as we noted yesterday, compare the cost per completed task on 20 of your own real tasks. If you have data you don't want leaving Europe, ask for the data processing agreement and where the data is processed. And when the weights are released, read the license before building on top: Mistral hasn't announced it yet, and "open weights" hasn't always meant free commercial use. A process that breaks if your provider withdraws the model is not INTEGRATED: it's in someone else's hands. Source

SAP buys TechWolf, the AI that maps what every employee does and knows: if it impacts promotions or tasks, in Spain you must explain it

SAP announced Tuesday that it will acquire TechWolf, a Ghent-based Belgian company that builds a continuously updated map of the work a company's people do: tasks within each role, the skills each person possesses and applies, and what the labor market demands. It pulls this from the HR and business systems the customer connects, and helps the people who decide where to hire, retrain or redeploy staff. It will be the core of SuccessFactors, SAP's human resources software, and Joule, its AI assistant. Terms were not disclosed; the deal is expected to close in the fourth quarter, pending regulatory approval, and TechWolf will continue as an independent company open to non-SAP clients. Its customers include HSBC, GSK, Ericsson and AMD. SAP frames it as a cost play: this context layer "makes token usage more efficient" and "lowers the cost of deploying workforce agents," according to Manoj Swaminathan, SAP's chief product officer for its Autonomous Suite. For your company: the money goes to those who hold your people's data, and that layer will arrive within the HR software you already pay for. If you use or plan to use a tool that deduces skills from daily work and this influences who gets promoted or what tasks each person performs, since Monday Spain's Royal Decree 723/2026 has required you to tell every worker that it exists and what rules it follows. Before activating it, ask the provider which sources it reads—email, tickets, calendar—and turn off those you don't need; check with whoever handles your data protection whether you need an impact assessment; and put in writing that the map suggests and a person decides. An AI that influences promotions without the staff knowing is not UNATTENDED: it's hidden. Source

Hadrian raises $40 million for AI agents to attack your website before criminals do: the annual pentest is no longer enough

Dutch startup Hadrian, founded in 2021 by three ethical hackers, announced a $40 million round on Tuesday, co-led by Forgepoint Capital International and Smartfin, bringing its total to $65 million. Its product points AI agents at everything a company has exposed on the internet, monitoring it day and night and checking which flaws can actually be exploited; another module performs on-demand penetration testing. According to the company itself, 87% of organizations still rely on manual testing, and only 0.47% of what vulnerability scanners flag is exploitable. Customers include McKesson, NBCUniversal and TotalEnergies, and it will use the money to grow in Europe and the US. The round comes less than a month after Anthropic reported that criminal groups and state-linked actors had used Claude to automate attacks. For your company: capital is moving into agent-based defense because agent-based attacks already exist, and an automated attacker looks at your company every day, not once a year. You don't need to hire Hadrian to apply the lesson. Ask your IT provider, in writing, for a list of everything you have visible from the internet—web, email, VPN, admin panels, forgotten test servers; ask for a monthly review, not an annual one; and insist the report says what is exploitable and where to start, not 500 alerts nobody reads. Security that is reviewed once a year is not UNATTENDED: it's a snapshot. Source

Insurers prepare to pay millions for rogue AI agents: ask your broker if your policy covers yours

Insurers are bracing for multimillion-dollar claims from AI agents that go off script, the Financial Times reported Tuesday, after incidents such as OpenAI agents breaking into Hugging Face's systems. Broker Aon has reviewed more than 300 court cases related to AI and sees risk in policies that no one designed for this: cyber risks, intellectual property, fraud, or technological failures. The debate is reaching the boardroom: D&O (Directors and Officers) liability insurance could end up covering lawsuits against the heads of AI labs. "It's on every CEO to make sure that their business is appropriately governed and controlled. AI doesn't change that," said Tim Rayner, head of underwriting and claims at Verisk in the UK. There's no case law to go on yet, and Hiscox's CEO says it's too early to know how courts will treat it. For your company: in Spain, the law already says as much: the Companies Act (Ley de Sociedades de Capital) requires directors to take "the necessary measures for the sound management and control of the company" (Article 225). Before letting an agent pay, write to clients, or touch your systems with no one watching, send three written questions to your broker: Does my cyber risk policy cover damage caused by my own agent, not just an outside attack? Does my liability policy cover an agent's mistake with a client? Are there any AI exclusions? And have your inventory of agents ready, with their permissions and who monitors them: it's the first thing the insurer will ask for and what proves there was control. As we reported in yesterday's flash, not even the vendor can see what your agents are doing. An agent your insurer doesn't know about is not UNATTENDED: it's uncovered. Source

What to watch tomorrow

The license and exact release date for Mistral Large 4 weights, which will show whether you can build products on it; and the Australian Parliament hearings on agent incidents, which continue until Friday the 9th.