Notes on data, AI, IT
and security
No marketing fog. The way I think about real problems with founders and managers.
A fifteen-person competence centre or an external architect: how a mid-sized plant covers the skills gap between production, IT and security
Large utilities are advised to build competence centres of 10 to 15 specialists at the junction of power engineering, IT and security. A plant of three hundred people cannot afford one and cannot hire one, yet it has the same critical infrastructure object, the same transition plan and the same model in the loop. What must stay inside, what can come from outside, and how to distribute signatures so that responsibility does not fall between four contractors.
Trusted hardware by 2030 and a machine on Windows XP: aligning a mandated transition with a line that runs until 2035
By 1 January 2030 significant critical infrastructure objects in Russia must predominantly run on trusted hardware and software platforms. Meanwhile the plant has a line on foreign controllers that will work for another ten years and cannot be stopped for a swap. What the decree actually requires, which nodes fall under it, where the no-equivalent exemption works, and how to turn compensating controls from an excuse into a justification.
Accident investigation needs a digital trace: what to capture before it happens
Since 1 January 2026 accidents and incidents in Russia's power sector are investigated under new rules, yet a security specialist is still not a mandatory member of the commission and a cyber attack is not listed as a cause. For a plant this means one thing: if configurations, firmware, logs and the model's state were not preserved at the moment of the accident, the investigation will close on the operator and the cause will stay in the loop. What to capture in advance, where to keep it and who signs.
FSTEC order 117 and AI in the loop: the regulator's first AI security checklist, translated into architecture
In April 2026 Russia's technical security regulator described AI protection in measures rather than terms for the first time: an isolated training environment, trusted data sets, model integrity, a log of every prompt and response, input and output filters, a ban on handing protected data to external services. Formally for state systems, in practice the first official checklist for any AI loop. Eight requirements translated into architecture and a one-quarter plan.
The air gap was an illusion before AI: what changes when models and agents live inside the loop
A piece on cyber resilience in the power sector dismantled the illusion of full isolation, split requirements into mission critical and business critical, flagged a gap in accident investigation and a shortage of cross-disciplinary staff. I agree with almost all of it and want to continue it: in 2026 the isolated loop has a new participant that makes decisions from data rather than code. That changes every one of his points.
Russia's 187-FZ nine years on: what changed since 2017 and what to do in 2026
In 2017 the critical infrastructure law was a frame, and a mid-sized plant could decide it did not apply. By 2026 self-assessment is closed by sector-wide object lists, trusted hardware is a plan with dates, foreign security tools are banned, and the regulator asks for running processes instead of binders. Four shifts, and the 2017 checklist rewritten for today.
When the provider is the incident: hybrid infrastructure and somebody else's maintenance
Azure had an infrastructure incident that hit hybrid scenarios, VPN and cloud VMware services. According to The Register, the trouble came out of internal maintenance work. Why this is the nastiest class of outage to own, and what actually helps.
Voice agents reached operations: what separates a working one from an expensive demo
ElevenLabs says its voice agents run more than 15 million conversations a week, and they now handle returns, insurance actions and bookings. Why voice became the first agentic AI segment with a real production use case, what separates a working agent from a polished demo, and what to settle about recordings and personal data before the pilot starts.
Observability for a swarm of AI agents: what they are doing, who allowed it, and how to stop them
OpenAI is building a dedicated mechanism to watch and control large numbers of concurrently running agents. Why agent observability differs from classic metrics, logs and traces, why it sits closer to security than to operations, and what to put in place while you still only have a handful of agents.
The coding agent gets its own environment: rights, review and the bill
OpenAI added reusable cloud development environments to Codex. A coding agent stops being autocomplete in an editor and becomes a participant in the pipeline with its own access. What changes in credentials, secrets, repository history, and who answers for the code it writes.
Kubernetes for AI agents: when always-on agents need a control plane
OpenClaw Enterprise is being built as an open-source control plane for always-on AI agents, with OpenAI, NVIDIA and Red Hat involved. What the AgentOps layer actually solves, who needs it today, who does not, and the signals that tell you it is time.
The FTC turns to autonomous agents: what you need to be able to prove about yours
A US regulator is asking whether existing law is enough to hold companies accountable for what autonomous AI systems do. While the legal frame is argued over, the engineering requirement is already clear: audit trail, permissions, sandboxing and human approval stop being optional.