
#CriticalOpsSecurity #Private5G #AIoT | #FutureOfLife | #SerialFounder | @VMO2Business | @GlobalTechAdv @TechLondonAdv | @UCL | @UniofBrighton Views are my own
London
Joined November 2009
- Tweets7.9K
- Following2.4K
- Followers2.3K
- Likes23.2K
Pinned Tweet
Securing critical infrastructure against AI-driven attacks can't rest on legacy firewalls and static IP rules. A defensible OT edge has three layers:
1. Standalone private 5G, isolated from public networks and corporate IT
2. On-site edge compute as the mandatory choke point
3. Context-aware, AI-driven inspection at that choke point
How are you securing the OT edge?
New: I've started Guilds, essays and a story about what gives life purpose once AI does more of the work.
Essay 1 starts with London's weavers in 1130.
guildsfutures.substack.com
2
1
144
Why I started Guilds, now an illustrated X Article: nitter.cf/Tom_Gere/status/212418…
4
Thomas Gere retweeted
The model provides capability. The harness sets the boundaries.
It governs:
→ what an agent may see
→ what it may do
→ when a human must approve
→ what gets recorded
The distinction becomes critical when AI moves from advice to action.
Nadella: "there is a lot of stuff in our critical infrastructure that's unpatchable."
Chillers, HVAC and transport, on the internet. Patching "instantly" into production is "going to create instability."
Contain OT on a private network with local edge apps. Monitor at the edge.
1
27
Source: TechCrunch techcrunch.com/2026/11/18/am…
10
This week's highlights from the AI CyberWatch independent monitor: attackers ran 25 autonomous AI operations against 19+ organisations, incl. chip makers. FBI: some AI-run intrusions reach privileged access in under 3 hours. Still protecting IT/OT: segmentation, private networks.
1
20
Sources: aicyberwatch.org/
Unit 42: unit42.paloaltonetworks.com/…
FBI FLASH via BleepingComputer: bleepingcomputer.com/news/se…
23
If AI agents are expected to deliver outcomes based on accurate information, what happens when they don't? Should the same rules apply as to a human colleague or a lower quality of service from a technical solution? In effect, should compensation and credits be due? It seems to be chaotic at the moment...
1
26
DeepMind says Apollo 2 humanoids on Gemini Robotics 2.5 kitted car parts on live Mercedes-Benz lines at 96% success.
Ars reports the model still runs in the cloud, not on the robot. In that setup, the network is part of the robot.
Robot, edge or cloud: where would you run it?
2
43
At 96%, @IlirAliu_ worked out that about 40 of every 1,000 picks still need a person. With @GoogleDeepMind's model running in the cloud for @Apptronik's Apollo 2, the plant network is part of that reliability budget. I'd measure link latency at the cell alongside pick success.
7
Source: Google DeepMind deepmind.google/blog/gemini-…
13
What does it take to let a robot be guided from off site?
Live video over a private 5G network the site runs. One controlled exit point, because remote guidance is remote access into OT, every shift. Each robot reaches only what its job needs.
#PhysicalAI #Robotics #Private5G
1
23
The original post, with sources: linkedin.com/posts/aboutthom…
4
Physical AI will reach plants "one production line, one control system and one verified decision at a time" (Fierce Network).
Stage 1: models in production. Stage 2: supervised agents. Stage 3: bounded control.
Stage 3 is where the network stops being plumbing. Your site?
1
21
Source: Fierce Network special report, 3 Nov fierce-network.com/cloud/spe…
1
12



