Can a language model help bring the power back after an attack
How large language models can help restore critical infrastructure after a cyberattack, and how far it is safe to trust them today.
The problem
Power plants, water supply, hospitals and telecommunications are connected to the internet today. After a cyberattack they have to get back to work quickly, but also safely: with the right order of steps, within the regulations and without bringing down something else that depends on them. The recovery plan is written under heavy pressure, and a mistake in it costs dearly.
Large language models look like a natural assistant for a task like this. Research on the topic, however, was scattered across cybersecurity, industrial control systems, digital twins and AI assurance, with no overall picture.
What we did
We reviewed the work on language models in incident response, in the resilience of industrial systems and in the recovery of cyber-physical systems. We arranged the approaches into five architectures, from the simplest to the most complex:
- an assistant that works with a prompt only;
- an assistant that searches the site’s documentation (retrieval-augmented);
- planning over a graph of the dependencies between systems;
- several agents that divide the work between them;
- hybrid systems in which the plan goes through verification or simulation before it reaches a human.
We mapped the real applications in energy, water, manufacturing, transport, healthcare and telecommunications, and grouped the limitations into four kinds: technical, security-related, governance-related and human-factor.
What it showed
On that basis we propose a reference architecture we call the Grounded Recovery Planning Stack, and a staged path from an assistant that helps a human towards bounded orchestration. The conclusion is that today the real value lies in assistants that are grounded in the data of the specific site, leave a trace for every decision and meet the requirements. Autonomous execution of recovery is premature until stronger verification, sector-specific test suites and formal safeguards appear.
Why then
The paper came out in June 2026, when language models were entering the day-to-day work of security teams faster than ways to trust them were appearing. It gives those teams a language for saying what is safe to automate today and what is not yet.
What comes next
The practical continuation of this work is CRIMS, a crisis management platform for municipalities, now in public beta.
The publication
Georgi Tsochev, Ivo Gergov. Large Language Models for Recovery Plan Generation in Internet-Connected Critical Infrastructures: Architectures, Applications, Limitations, and Research Directions. Future Internet 18, 295 (MDPI, 2026). doi:10.3390/fi18060295