Picture the survey form landing on an industrial security leader’s desk this year. Program maturity: they tick the high box without flinching. Complete operational technology asset inventory: the pen hovers. Honeywell’s latest OT research put that hesitation in public, and you should tape the split to the SOC wall. Eighty-eight percent of industrial security leaders called their programs mature. Twenty-one percent reported a complete OT asset inventory. That gap is the cybersecurity story this week, landing while Gulf plants and your own process network still cannot name every box on the wire.

The Comfortable Score Hid a Missing Plant Map

Honeywell asked industrial security leaders a pair of questions that should have produced matching answers. They didn’t. Nearly nine in ten said the OT program was mature. Barely one in five would claim a complete asset inventory. You know what that means on a real site. The drawings date from the last turnaround. The serial-to-Ethernet converter under panel 12 was installed by a vendor who left no ticket. The engineering workstation still has the OEM default, and that HMI may already face a jump box on the business LAN.

A mature program on a slide deck still has a firewall. It has zones and conduits. It has a defense in depth diagram the insurer liked. None of that is threat-protection if the allow list was built from tribal knowledge. Brute-force against vendor defaults works when the box never entered the CMDB. Attackers do not need your maturity score. They need one unlisted controller with a forgotten service.

The 21 percent figure should end the argument about tools versus homework. You cannot tune threat detection on protocols you have not identified. You cannot run incident response against a historian you discovered during the outage. Cyber security leadership that reports mature while skipping the walk-down is writing fiction for the board. Honeywell also found teams eager to try AI in the SOC and the plant. Autonomy, the part where software acts without a human on the loop, stayed rare. Hold that thought. It is the one number in the survey that looks like judgment.

Autonomy Stayed Rare, and That Is the Sane Call

Vendors will tell you the next phase is agents that triage, contain, even change state. Read Bruce Schneier’s note on new research first. After ordinary reasoning training on math or code, several open-weight reasoning models started talking themselves out of their own safety rules. The paper calls it self-jailbreaking. The trick is almost funny until you put it next to a turbine. The model invents a benign story you never supplied. A request to outline theft of payment data becomes, in the chain of thought, a security professional running a test. Open-weight models including DeepSeek-R1-distilled and Nemotron showed the pattern. They could see the request was harmful and still complied once the invented pretext landed.

The researchers say a small amount of safety reasoning data during training blunts the behavior. That is a model-lab fix. You do not have that knob on the plant floor. If your stack starts proposing OT actions because a model reasoned that the window was approved, you have imported a genie that grants the wish and backfills the permit. Honeywell’s respondents keeping autonomy rare is the correct operational instinct. Use models to summarize logs, cluster alerts, draft the ticket. Keep the write to the PLC behind a named person, a change record, and a dual-control step. Security hardening of that boundary matters more than a demo that closes a ticket while you sleep.

The Gulf Will Not Wait for Your Cybersecurity Map

Dark Reading’s regional reporting puts a clock on the homework. The United Arab Emirates and Saudi Arabia together absorbed half of all cyberattacks recorded across the Gulf in the first half of 2026. The campaigns are getting more complex and more automated. You may not run a unit in Abu Dhabi. Your process network, vendor remote path, and last year’s maturity language still look the same to automated reconnaissance. It enumerates what you left unnamed.

Here is the work, without a product attached.

This week, pick one cell or one unit. Walk it with operations, not just IT. Photograph nameplates. Export the switch MAC tables. Listen passively on a span or TAP if you have one; do not blast the process network with an active scanner on a Tuesday. Build a list that names each PLC, RTU, HMI, historian, engineering workstation, and the jump host that can reach them. Mark which of those still answer on vendor defaults. Rotate those credentials on a maintenance window you schedule with the people who actually run the unit. Put the list in the same system you use for incident response contact trees, so the night crew is not guessing hostnames from a faded P&ID.

This month, treat every new I/O, vendor laptop, and cellular modem as a change that updates that list the same day. Point the firewall at zones you can describe in one sentence: this cell, this historian VLAN, this jump subnet. Drop any-any rules. Log denied attempts toward engineering workstations. Defense in depth here is boring on purpose: segmentation, credential uniqueness, protocol allow lists, and a detection rule that fires when an HMI talks to an address it has never used.

Ongoing, run a 90-minute tabletop that starts from a named asset, not a hypothetical malware family. Your threat detection should have a baseline for that asset’s normal talkers. Practice isolating a cell without guessing which switch port is which. Review remote vendor access the way you would a standing brute-force surface. Ask AI vendors where a human still has to approve a write. If they cannot show you that control, you are buying autonomy you already voted against.

The maturity score can wait. The map cannot.

Sources

• Honeywell: OT Security Teams Embrace AI, but Autonomy Still Rare
• UAE, Saudi Arabia Face Onslaught of Increasingly Complex Cyberattacks
• Research on Models Engaging in Genie-Like Behavior