NB593: AI Agents Come for Your Network; Google to Send TPUs Into Orbit

Drew Connery-Mauri (0:00)

Take a network break, I'm Drew Connery-Mauri.

Jonna Johnson (0:02)

I'm Jonna Johnson.

Drew Connery-Mauri (0:03)

And today we're going to talk about Cisco research into agentic AI adoption and network operations, new agentic AI capabilities in Selector and NetBrain, China taking a hard look at its dependence on Broadcom, financial results from the quantum realm, and some interesting space networking news. But first, a message from our sponsor, Check Point. Today's sponsor is Check Point. Traditional firewalls were not designed for securing AI traffic. They remain completely blind to the meaning inside the modern AI layer, unable to parse AI prompts sent in any language or govern MCP streams and autonomous agents. The network is the ideal place to see and control it in real time. Check Point's AI network firewall builds this semantic intelligence natively into your existing footprint.

No new hardware to rack, no separate management console to learn. It applies identically across on-premises and cloud deployments. Check Point has secured the most demanding enterprises and data centers for over 30 years. Learn more at checkpoint.com and do a quick search on AI network firewall. We thank Check Point for being a sponsor. We also have a Tech Bytes show after the news. Sponsor Broadcom is here to make the case that private cloud is the place for AI inferencing workloads. I talked to Broadcom about the new VMware AI factory. It's combining software and hardware to accelerate deployment and operations, curated models that make it easy to deliver inferencing as a service to internal customers and more.

Before the news, John, let's hit the red alert.

All right.

Jonna Johnson (1:25)

In the past seven days, there were 2,970 CVEs created with 7,204 updated. Also, 10 known exploited vulnerabilities in the past week with 47 in the past month. Speaking of exploits, Check Point has said that active exploits have been launched against its security gateway and our red alert in fact goes to them. The company is warning of active exploits against two known vulnerabilities, that is CVE 202685102, which with a CVSS of 9.8.

Notably, these have both been patched, so you know where we're going with this, apply the patches.

Drew Connery-Mauri (2:04)

Yeah.

Jonna Johnson (2:05)

Drew, want to talk about what some of these are?

Drew Connery-Mauri (2:08)

Yeah. So the first exploit is taking advantage of remote code execution vulnerability and how Check Point security gateway handles VPN certs. The second exploit is targeting a pre-authentication path traversal flow in the management web interface. And as you mentioned, these vulnerabilities are already known, Check Point did address them. They also have workarounds if you can't get to the patches.

But they are now saying active exploits are in the wild. So if you are on this platform, it's time to dive in.

Jonna Johnson (2:34)

Yeah. And it's worth tracking the rate of increase of active exploits as well as the development of CVEs. It's really quite interesting if you look at the trend overall, because starting the beginning of the summer, the rate of increase has gone up dramatically. So I guess that's what second derivative.

Drew Connery-Mauri (2:51)

Yeah.

Jonna Johnson (2:51)

And in any event, it is coincides. I don't think this is by chance. It coincides with Mythos coming out. So very interesting to see the impact of Mythos.

Drew Connery-Mauri (3:02)

Yeah, for sure.

All right. Let's dive into some news. We'll lead off with Cisco. They released a report looking at how network operators are using AI agents. Cisco partnered with Omnia to survey a thousand respondents in IT and networking. The report says 51 percent are running agentic AI that acts not just advises in production networks. The report also says organizations are deploying agentic AI to keep up with alerts and events. An average of about 2,100 per day among respondents just because they don't have enough human operators.

This was surprising to me. Fifty-six percent of respondents allow agentic AI to make changes with human approval. Another 24 percent allow full autonomous changes. Although I will note with the caveat for some processes including, quote, rerouting traffic, adjusting wireless parameters, isolating suspicious endpoints and resolving incidents end-to-end. Jonna, your thoughts?

Jonna Johnson (3:52)

Well, I think the main thing to keep an eye on is surveys are very, very difficult beasts. I know we've done surveys in the past. We prefer to do interviews for a very good reason because you get to go deeper in what's actually going on. Part of the challenge is it comes down to how the question is worded. So if you say something like, do you have any processes inside your entire environment that are allowed to run without human approval?

You may have 9,999 processes that require human approval, but one that doesn't, you're going to check the boxes as we do. And then what will happen is you'll get a pop-up that says, oh, describe what kind of process that is. So it's a little bit tricky. I'm not arguing with the data. I'm just saying that the data can be presented in one way versus another. And without more insight, it's hard to say.