Skip to content
Hakan KaracaENGINEERING JOURNAL
AI & Intelligent AutomationEki 8, 20264 min read

AI-Assisted OT Cybersecurity: Useful Detection, Dangerous Automation

AI can reduce the volume of security information an analyst must read, but OT actions still need controlled procedures.

AI can help security teams sift through large volumes of alerts, correlate asset information and summarise incident evidence. In operational technology, however, the cost of an erroneous action may include production interruption, loss of visibility or a safety consequence. The design objective should be better analyst decisions, not maximum automation for its own sake.

Why OT changes the security problem

Industrial networks operate under availability, safety, reliability and timing constraints that differ from ordinary office IT. A device may run a legacy operating system because the validated equipment configuration depends on it. A maintenance connection may be essential during an outage but inappropriate during production. A sudden network scan could be merely noisy on a laptop fleet and operationally disruptive on a fragile controller network.

Security analytics must therefore understand asset role, communications patterns, maintenance windows, process state and the consequence of a proposed response. An alert score without asset context can generate a long queue of distractions. At 03:00, the on-call engineer does not need 400 red indicators that all mean “something might be interesting”.

Good use cases for AI assistance

  • Alert triage: group related events and provide a summary with original log references and timestamps.
  • Asset and vulnerability context: help reconcile inventory records, software versions and known advisories while showing source and freshness.
  • Investigation support: compare an observed sequence with documented normal behaviour and suggest evidence to collect.
  • Incident documentation: draft timelines or handover notes from approved records, subject to analyst review.
  • Threat-hunting queries: assist a qualified analyst in formulating queries for logs or asset inventories before execution in an authorised environment.

For these tasks, the model can reduce search and writing effort without becoming the authority for a production action.

Where automatic responses become hazardous

Blocking a workstation, disabling an account, quarantining an asset or changing a firewall rule can be reasonable in some IT contexts. In OT, those actions should be chosen through an approved response plan that accounts for the system’s function and operational state. A generic model must not decide to isolate a controller because an unfamiliar traffic pattern resembles an attack; the pattern may be a legitimate engineering download or a recovery process.

Use graduated response levels. Low-risk actions might enrich an alert or open a case. Higher-impact changes require explicit authorisation, pre-approved playbooks, asset-aware checks and an auditable approval step. Emergency procedures should be tested with operations, engineering, IT and safety stakeholders—not invented during the incident.

Model security is part of OT security

Prompts, retrieved files, log messages and ticket descriptions are data, not privileged instructions. A malicious string embedded in a document should not be able to change an agent’s permissions or persuade it to expose secrets. Restrict tools to narrowly scoped operations, validate parameters, require server-side authorisation and separate the model from privileged credentials.

Network segmentation remains important. An AI service that can query the internet, browse a document repository and write configuration changes to a control network has accumulated several different trust domains. Keep the interfaces narrow, log access, limit outbound connections and make the service’s failure mode explicit. A model outage should not remove the operator’s ability to use approved procedures or manage the process.

Build an evaluation set from real operations

Test known benign events, real incidents where available, uncommon but legitimate maintenance activities, missing asset records, stale vulnerability feeds and deliberately misleading text. Ask whether the system can say that evidence is insufficient. Measure false positives, missed detections, source correctness, analyst time and the safety of recommended actions.

Run the system in observe-only mode before enabling any response. Compare its recommendations with analyst decisions, review disagreements and identify which failures arise from data quality, asset context, model reasoning or poor workflow design. A confidence score is not a risk assessment.

Governance and evidence

Document the use case, system owner, data sources, access model, acceptable actions, human approval requirements, performance limits and rollback plan. Version the model and prompts, preserve relevant inputs and outputs according to the site’s security policy, and periodically reassess the service after changes to assets, network topology or threat intelligence.

Recommendation

Use AI to help experts find patterns and evidence faster. Keep containment and control actions bound to pre-approved, risk-assessed procedures with explicit authority. The right measure of success is not how many decisions the model makes; it is whether the security team resolves meaningful incidents faster without creating a second incident in the process.

References and further reading

  1. NIST SP 800-82 Rev. 3: Guide to Operational Technology Security
  2. CISA: Industrial Control Systems Recommended Practices
  3. NIST AI Risk Management Framework