The Attackers Moved Faster Than The Rules

Start with the number, because it is the only honest place to start. Only 13% of OT network segments are fully isolated, according to the SecurityWeek analysis published this week. That is not a statistic about a lagging industry sector, it is a description of how most industrial networks actually work: as one continuous surface, where a foothold in one place is a foothold almost everywhere else. Everything else in this week's reporting, from a UK official's warning to a fresh CVSS 10.0 flaw under active exploitation, sits on top of that fact.

The 13% that explains everything

Segmentation is not a nice-to-have. It is the thing that turns a single compromised device into a contained incident rather than a total one. When 13% is the ceiling for full isolation, the remaining OT segments are where an intruder who gets in once can keep going, often without needing a second exploit at all.

This week supplied plenty of first exploits to worry about. SecurityWeek reported that a recent ZyXEL switch vulnerability is being exploited by Chinese hackers. BleepingComputer flagged a maximum severity zero-day in D-Link DIR-822A routers. The Hacker News covered a CVSS 10.0 flaw in VeloCloud Orchestrator, actively exploited specifically in certificate-based setups. None of these are exotic. They are the ordinary hardware sitting at the edge of ordinary networks, and each one is a door. The 13% figure tells you what happens once someone walks through it: not much resistance.

More than a third of industrial organisations now see cybersecurity risk as a top obstacle to growth, according to the Dark Reading study out this week. Read that alongside the segmentation figure and the causality runs the wrong way for comfort. It is not that these organisations are cautious because they are exposed. It is that they are exposed, know it, and have not fixed the underlying architecture, so the caution never converts into growth or into safety.

AI agents don't wait for permission

The Hacker News' piece on AI agents rewriting the rules of lateral movement is the sharpest single argument in this week's coverage, because it names the mechanism directly. Lateral movement used to require a human attacker to make decisions: which credential to try, which share to check, which machine looked interesting. That decision-making was slow, and slowness was the defender's only real advantage. An analyst reviewing logs had time, because the attacker had to think between steps.

Agents remove that thinking-time. They enumerate, decide and move without a human pausing to consider the next action, so the gap between initial access and a wider compromise gets shorter. In a network where only 13% of OT segments are fully isolated, that speed has almost nothing to push against. The fences that would have made an agent pause, or made its movement visible as it crossed a boundary, are largely not there. An attacker moving faster is a problem. An attacker moving faster through a network with no internal walls is a different problem, and it is the one this week's reporting keeps circling back to.

This is also the week Schneier on Security reported that GPT-6 Astra broke an old Enigma message. Set aside what that says about cryptography specifically; what it says about capability is the point worth keeping. Systems are now solving problems that used to demand sustained, specialised human effort, and treating it as a side note. Nightmare Eclipse dropping a new Microsoft Defender exploit after revealing its own identity, and a SharePoint flaw that Microsoft initially listed as spoofing turning out to enable authenticated remote code execution, both reported this week, are further evidence that the offensive side is not short of new material. A Linux kernel flaw giving ARM64 KVM guests read-write access to host memory adds one more. None of these needed AI to be found. What AI changes is how fast they get chained together once they are.

Frameworks built for slower enemies

The Hacker News asked, in its DORA Year Two coverage, whether a security operations centre can actually see the attack. It is the right question, and this week's reporting suggests the honest answer for a lot of organisations is no, not reliably, and not fast enough. DORA and frameworks like it were written on an assumption: that an intrusion unfolds over a timeframe that gives monitoring and response a fair chance to intervene. That assumption held reasonably well against human-paced attackers. It holds much less well against the kind of lateral movement The Hacker News describes AI agents now performing.

Visibility depends on structure. A SOC can only see an attack in any meaningful sense if the network is divided so that movement between segments becomes a detectable event rather than an invisible default. With only 13% of OT segments fully isolated, most movement inside a compromised industrial network trips nothing at all. It just looks like normal traffic, on a network that was never cut up to make it look like anything else.

The same visibility problem is not confined to industrial settings. BleepingComputer is running a webinar this week on real-world Google Workspace breaches, which points at the same gap from the enterprise side: organisations discovering, after the fact, what an attacker actually did inside systems they assumed were being watched. And it is not only network flaws doing the damage. SecurityWeek reported this week that a malicious B-tree NPM package has accumulated millions of downloads, which is a supply-chain route into exactly the same unsegmented environments, no zero-day required.

What the official actually said

It is worth taking the UK official's comment to The Record at face value rather than as a hedge. AI is set to help attackers much more than defenders, they said, and this week's reporting does not offer much to argue against that. Defenders' gains from AI this week are mostly prospective: better log analysis, better alert triage, the promise of a SOC that finally keeps up. Attackers' gains are already operational: agents performing lateral movement, exploits landing against maximum-severity flaws within days of disclosure, a language model breaking a decades-old cipher message as an aside.

None of that is an argument against defensive AI. It is an argument that the asymmetry is real and current, not theoretical, and that it is compounding on top of an architectural failure that predates AI entirely. Only 13% of OT segments are fully isolated because segmentation is slow, expensive and unglamorous work that keeps losing out to other priorities. AI did not create that gap. It just found it, and it is moving through it faster than the frameworks written to catch it were ever designed to cope with.