Skip to content

Runtime: Microsoft wants to be the safe house for enterprise AI

+ more details on the Microsoft 365 debacle from last week, how companies are dealing with personal AI agents, and more.

Runtime: Microsoft wants to be the safe house for enterprise AI
Microsoft Cobalt servers. (Credit: Microsoft)

Welcome to Runtime! Today: Microsoft lays out its pitch to the enterprise at Ignite while shedding a little more light on the bug that took down Outlook last week, and more.

Please forward this email to a friend or colleague! If it was forwarded to you, sign up here to get Runtime for free every week, or level up here.


First up: Model behavior

SAN FRANCISCO — Microsoft's first-mover advantage in enterprise AI eroded about a year ago after a new batch of Anthropic's models paved the way for the AI coding boom, forcing longtime partner OpenAI to quickly pivot to enterprise product development and Microsoft to embrace Anthropic. But it's clear that the agentic AI era will force companies to learn new ways to build a business on the internet, and on Tuesday Microsoft outlined several new products and services that aim to give customers a safe way to operate in this new world.

CEO Satya Nadella announced seven new AI models and the general availability of a few key services announced earlier this year, such as Microsoft Perception and Microsoft IQ. "AI in the flow of human ambition is perhaps the biggest design challenge we have in front of us," Nadella said in his keynote address, emphasizing that for all the talk of the autonomous enterprise, most companies still want to make sure someone is watching.

Three of the new models were designed for cybersecurity and coding tasks, and Microsoft also announced the general availability of Decision 1.1, its first decision-model family. Nadella singled out the growing interest in decision models: "Just like how databases back in the day or relational databases back in the day changed software engineering, I think of decision engines similarly," he said.

Microsoft also sought to strike a balance between the opacity of closed weight models from Anthropic and OpenAI and the free-wheeling availability of open-weight models like GLM or Kimi. It will allow companies that are part of its Frontier Company consulting engagement to adopt MAI Licensed Weights, which allows companies to "adapt them to your business and safely integrate your proprietary knowledge and expertise," it said in a blog post.

Finally, "at the end of the day, we believe there should never be a choice between innovation and safety," Nadella said, drawing yet another clear distinction between Microsoft and the AI labs. He announced the AI Transparency Initiative as a bid to give AI customers a way to track what happens as AI models do the work between the prompt and the result, but without the participation of the model labs or its fellow hyperscalers, it's hard to see that catching on.


The rest of The Stack

Bot battles: Personal AI agents aren't showing up in enterprise customer-service call logs as much as you might think given the enthusiasm of their backers, but they are posing new problems and challenges for businesses that designed their customer feedback systems around people. Technical solutions like the Personal Agent Protocol could help, and some companies might learn that they prefer to work with personal agents given they don't have to spend expensive tokens trying to get the linguistics right.

Personal AI agents are calling customer service. No one is sure how to handle them
Personal AI agents are not widely used, but they're very different: “Halfway through the call, the AI agent starts crying.”

Rule of three: Nadella apologized for last week's epic Microsoft 365 outage at the top of his keynote, which was "caused by three main factors that together caused a global outage for the networking gateway that fronts many Microsoft 365 services," according to Microsoft's Charles Lamana. The progression: A July networking update introduced a "latent bug" that was triggered by a separate update last week, everything failed over onto a single network that quickly became overwhelmed, and much-needed extra capacity that could have helped was offline for maintenance.

Microsoft says latent bug knocked 80% of its networking servers offline
Load testing had focused on "weaker hardware SKUs with fewer cores”

Airheads: The FAA has still not fully explained the networking error that caused Friday's hours-long ground stop, throwing the US air travel system into chaos, but they're blaming it on a "legacy" network that was in the process of being upgraded. A faulty router also popped up in several interviews with Secretary of Transportation Sean Duffy, but the government has yet to provide specific details on how that could have happened.

FAA blames “legacy TDM” network for flight-grounding
x25 protocol and ancient copper cables…

It's a trap: The open-source ARTEX hacking tool has been the vehicle for several AI hacking attempts over the last month, and Amazon managed to catch several groups attempting to steal credentials for AI services. A network set up to look at AI inference APIs harvested a lot of data: "Each call shipped system prompts, task context, tool-use histories, knowledge-graph state  agents' reported findings straight into Amazon's telemetry," according to Amazon CISO CJ Moses.

Amazon spins up AI honeypot, grabs ARTEX attack chain.
"System prompts, task context, tool-use histories, knowledge-graph state... straight into Amazon's telemetry."

Stack ranking

Brian Halligan is once again the CEO of HubSpot, returning to the marketing company he co-founded after a little over five years since he stepped down.

Steve Kelley is the new chairman and CEO of Expel, joining the MDR security company after executive leadership roles at Bitdefender and EdgeWave.


We're also reading:

GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold: Google threat researchers said they've identified a milestone: hacking groups are using AI tools to manage agent swarms autonomously.

Feeling Easier, Working Harder: Cognitive Load in AI-Assisted Code Comprehension: "This study investigates how access to generative AI affects developers' real-time cognitive workload, code comprehension performance, and self-perceived code comprehension strategies," the researchers wrote, and a year after the explosion of AI coding agents feels like a good time to take stock of that situation.

Updating our coordinated vulnerability disclosure policy timelines: Anthropic is reducing the disclosure window "from 90 days to 30 days for high- and critical-severity issues, and to 45 days for all other issues," and that is going to cause a lot of frustration across open-source circles and small businesses.


Thanks for reading — see you Thursday!

Add The Stack on Google