Microsoft Cautiously Onboards Grok 4 to Azure AI Foundry Amid Safety Concerns

Microsoft is cautiously onboarding Grok 4 following Hitler concerns

Earlier this year, Microsoft CEO Satya Nadella moved with haste to get engineers to test and deploy DeepSeek's R1 model on Azure AI Foundry. It was an unusually quick turnaround for a new model on Azure that also set a new bar for success. A few months later, Nadella then pushed to onboard xAI's Grok 3 models, in a deal that saw them arrive on Azure AI Foundry just in time for the first day of Microsoft's Build developer conference in May. Elon Musk even appeared during Nadella's Build keynote, in a somewhat jovial conversation about his early days as a Microsoft intern -- despite Musk making Microsoft a defendant in his lawsuit against OpenAI. I'm now hearing from sources familiar with Microsoft's AI plans that the company is taking a far more careful approach with onboarding xAI's latest Grok 4 model, after quickly onboarding models from competitors like OpenAI, Meta, and Mistral. Musk introduced Grok 4 early last month, just days after the Grok chatbot spewed out a series of pro-Hitler views on X. I'm told that Grok's Nazi sympathizing set off alarm bells inside Microsoft, just as the company was preparing to launch Grok 4 on Azure AI Foundry. Instead of simultaneously shipping Grok 4 on Azure, like Microsoft does with most of OpenAI's new models, there has been no public announcement of when xAI's new model will be available through Microsoft's AI services. I understand Microsoft has been red teaming Grok 4 throughout July, a process where teams attack AI systems to find vulnerabilities or safety issues. One source described some of the early red team reports for Grok 4 as "very ugly." Instead of launching Grok 4 to all Azure AI Foundry customers, Microsoft is now working on a private preview of the model -- just as Grok hits the headlines again for generating Taylor Swift nudes. This preview will only be available for a handful of customers, I'm told, as Microsoft wants to ensure Grok 4 is enterprise ready. I asked Microsoft to comment on its Grok 4 plans for Azure AI Foundry, but the company refused to provide a statement that's attributable to a named spokesperson. Grok 4 availability on Azure AI Foundry is important to xAI, as it gives Elon Musk's AI company direct access to Microsoft's enterprise customers. It's equally important to Microsoft, as the company is keen to position itself as the AI model hosting company. Microsoft's circumspect approach with the Grok 4 preview means it's unlikely that we'll see xAI's latest model appear more broadly on Azure AI Foundry anytime soon. Microsoft announced another AI shake-up this week, affecting Charles Lamanna's Business & Industry Copilot (BIC) teams. In June, the BIC team moved closer to the Microsoft 365 Copilot side of the company, with Lamanna reporting up to Rajesh Jha, Microsoft's head of experiences and devices. In an internal memo yesterday, Lamanna announced some BIC leadership changes that will see Microsoft make Agent 365 an "official product initiative." Agent 365 sounds like a way for Microsoft to focus on the challenges of AI agent security and compliance while it scales these agents across Teams, Outlook, and SharePoint. The Agent 365 initiative will be led by Nirav Shah, a 24-year veteran at Microsoft. Microsoft is also combining parts of its Power Automate and Copilot Studio teams together. "Agent flows and CUA from Power Automate will move to Copilot Studio under Dan Lewis," says Lamanna in his memo. Some other Power Automate employees will now be part of Microsoft's Power Platform team. Finally, the BIC team is also creating "Microsoft Forward Deployed Engineers (FDEs)." These are technical experts who work directly with Microsoft's customers to get them to activate AI products. The FDE announcement comes amid Microsoft's layoffs that have targeted sales teams as part of a broader restructuring effort to bring in more technical salespeople to better demonstrate AI tools to businesses. "Forward Deployed Engineers are quickly becoming essential to large scale AI transformations, as seen at companies like Plantir and OpenAI," says Lamanna. "Just as we are changing the way we do software and product development, Microsoft will need to adopt this new way of delivering AI solutions to customers." I'm always keen to hear from readers, so please drop a comment here, or you can reach me at [email protected] if you want to discuss anything else. If you've heard about any of Microsoft's secret projects, you can reach me via email at [email protected] or speak to me confidentially on the Signal messaging app, where I'm tomwarren.01. I'm also tomwarren on Telegram, if you'd prefer to chat there. Thanks for subscribing to Notepad.

Microsoft tests Grok 4 in restricted Azure preview

According to The Verge, Microsoft is proceeding deliberately with the onboarding of xAI's Grok 4 model onto its Azure AI Foundry, reflecting a more cautious approach compared to previous rapid deployments of other AI models. Earlier this year, Satya Nadella, Microsoft CEO, expedited the testing and deployment of DeepSeek's R1 model on Azure AI Foundry. This process marked an unusually quick turnaround for a new model on Azure, establishing a new benchmark for efficiency. A few months subsequent to the DeepSeek R1 deployment, Nadella initiated a push to onboard xAI's Grok 3 models. This arrangement ensured the models' availability on Azure AI Foundry by the commencement of Microsoft's Build developer conference in May. During the Build keynote, Elon Musk made an appearance alongside Nadella. Their conversation touched upon Musk's early tenure as a Microsoft intern, despite Musk having named Microsoft as a defendant in his lawsuit against OpenAI. Microsoft is now adopting a significantly more careful strategy for integrating xAI's latest offering, the Grok 4 model. This cautious stance follows the company's prior swift onboarding of models from various competitors, including OpenAI, Meta, and Mistral. Elon Musk introduced Grok 4 early last month. This introduction occurred just days after the Grok chatbot generated a series of pro-Hitler views on X. Reports indicate that Grok's generation of content sympathetic to Nazi ideology caused considerable concern within Microsoft, particularly as the company was in preparations to launch Grok 4 on Azure AI Foundry. In contrast to Microsoft's typical practice of simultaneously shipping new OpenAI models on Azure, there has been no public announcement regarding the availability of xAI's new model through Microsoft's AI services. Microsoft has engaged in red teaming of Grok 4 throughout July. This process involves teams actively probing AI systems to identify vulnerabilities or safety concerns. One source characterized some of the initial red team reports for Grok 4 as "very ugly." Instead of deploying Grok 4 to all Azure AI Foundry customers, Microsoft is currently developing a private preview of the model. This development coincides with Grok once again attracting attention, this time for generating unauthorized images of Taylor Swift. This private preview will be accessible only to a limited number of customers. Microsoft's objective is to ensure that Grok 4 is ready for enterprise-level application. When questioned about its plans for Grok 4 on Azure AI Foundry, Microsoft declined to provide an attributable statement from a named spokesperson. The availability of Grok 4 on Azure AI Foundry holds importance for xAI, as it would grant Elon Musk's AI company direct access to Microsoft's enterprise customer base. This availability is equally significant for Microsoft, which aims to position itself as a premier host for AI models. Microsoft's measured approach with the Grok 4 private preview suggests that xAI's latest model is unlikely to become broadly available on Azure AI Foundry in the near future. Microsoft recently announced another reorganization within its AI division, impacting the Business & Industry Copilot (BIC) teams, which fall under Charles Lamanna. In June, the BIC team was integrated more closely with the Microsoft 365 Copilot side of the company, with Lamanna now reporting to Rajesh Jha, who leads Microsoft's experiences and devices division. In an internal memorandum issued yesterday, Lamanna outlined several leadership adjustments within BIC. These changes include Microsoft's designation of Agent 365 as an "official product initiative." Agent 365 appears to represent Microsoft's strategy to address the complexities of AI agent security and compliance while concurrently expanding these agents across Teams, Outlook, and SharePoint. This Agent 365 initiative will be overseen by Nirav Shah, who has a 24-year tenure at Microsoft. Microsoft is also in the process of merging specific components of its Power Automate and Copilot Studio teams. Lamanna stated in his memo that "Agent flows and CUA from Power Automate will move to Copilot Studio under Dan Lewis." Other employees from Power Automate will be reassigned to Microsoft's Power Platform team. Additionally, the BIC team is establishing "Microsoft Forward Deployed Engineers (FDEs)." These individuals are technical experts tasked with collaborating directly with Microsoft's customers to facilitate the activation of AI products. The announcement regarding FDEs comes amidst Microsoft's ongoing layoffs, which have specifically targeted sales teams. These layoffs are part of a broader restructuring effort aimed at recruiting more technical sales personnel to enhance the demonstration of AI tools to businesses. Lamanna stated, "Forward Deployed Engineers are quickly becoming essential to large scale AI transformations, as seen at companies like Plantir and OpenAI." He further added, "Just as we are changing the way we do software and product development, Microsoft will need to adopt this new way of delivering AI solutions to customers."