CISA, NSA and Partners Publish AI-in-OT Security Principles

CISA and the NSA, together with Australia's ACSC and cyber agencies from Canada, Germany, the Netherlands, New Zealand, the United Kingdom and the FBI, published joint guidance on securely integrating artificial intelligence into operational technology environments.

The guidance, Principles for the Secure Integration of Artificial Intelligence in Operational Technology, was published on December 3, 2025.

What it is

The document is a joint cybersecurity information sheet that outlines four principles to help critical infrastructure owners and operators integrate AI into OT systems while managing new risks to safety, security and reliability.

It focuses on machine learning and large language model-based AI and AI agents because integrating OT with these types of systems involves more complex safety and security considerations.

The authoring agencies encourage critical infrastructure owners and operators to review the guidance and act on its principles so they can safely and securely integrate AI into OT systems.

What it requires

The guidance sets out four steps, starting with a call to understand AI: educate personnel on AI risks, impacts, and secure development lifecycles.

The second step tells operators to assess AI use in OT by evaluating business cases, managing OT data security risks, and addressing immediate and long-term integration challenges.

The third step calls for establishing AI governance by implementing governance frameworks, testing AI models continuously, and ensuring regulatory compliance.

The fourth step directs organizations to embed safety and security by maintaining oversight, ensuring transparency, and integrating AI into incident response plans.

Supporting mitigations in the guidance include pushing data from the OT environment to a separate AI system where appropriate and implementing fail-safe mechanisms to limit the consequences of failures and worst-case scenarios.

What security leaders should do

Security leaders overseeing OT environments should map where AI, machine learning or AI agents touch Purdue-model levels zero through five before approving any deployment, since the guidance treats that mapping as the starting point for risk assessment. Governance teams should pair OT engineering and security functions when evaluating vendor AI features, given the guidance's emphasis on understanding how vendors handle OT data. CISOs should also confirm that monitoring, human-in-the-loop checkpoints and fail-safe mechanisms are documented in incident response plans before AI systems are connected to OT networks, and should treat this joint guidance as a baseline checklist when briefing boards or regulators on AI-in-OT risk management, since it carries the backing of CISA, NSA and seven partner cyber agencies.