Full Report
The guidance gives operators a clearer map, and it reinforces that resilience grows when humans and machines work in partnership. The post New cybersecurity guidance paves the way for AI in critical infrastructure appeared first on CyberScoop.
Analysis Summary
# Main Topic
New unified cybersecurity guidance issued by global agencies (CISA, FBI, NSA, ASD/ACSC, and partners) detailing secure integration principles for Artificial Intelligence (AI) within Critical Infrastructure (CI) Operational Technology (OT) environments. The guidance aims to provide practical guardrails, emphasizing human-machine partnership for resilience while managing significant AI-related risks.
## Key Points
- **Shift to Practicality:** The guidance moves beyond theoretical debate to offer practical direction for AI deployment in OT environments, notably through the "Principles for the Secure Integration of Artificial Intelligence in Operational Technology."
- **Safety vs. Security Distinction:** A critical finding is the explicit separation between safety (preventing physical harm) and security (protecting data/availability).
- **LLM Limitation:** Large Language Models (LLMs) are explicitly advised **not** to be used for making safety decisions in OT environments due to AI's non-deterministic nature and risk of hallucinations.
- **AI Role Definition:** AI should function as an "adviser rather than a controller." Predictive Machine Learning is suitable for Levels 0-3 (e.g., failure forecasting), while LLMs are better suited for Levels 4-5 (business/documentation functions).
- **Human Element Risk:** Heavy reliance on AI may cause OT personnel to lose necessary manual skills required for managing systems during failures.
## Threat Actors
- No specific named threat actors or campaign attribution were detailed in relation to the issuance or context of this guidance; the focus is on establishing foundational security principles against inherent technological risks.
## TTPs
- The guidance focuses on mitigating risks introduced by AI itself, rather than cataloging adversary TTPs. Key risks highlighted include:
- OT process models "drifting over time."
- Safety-process bypasses enabled by AI.
- Unpredictable behaviors or hallucinations from modern AI models (like LLMs).
## Affected Systems
- **Critical Infrastructure Operational Technology (OT):** Systems spanning Levels 0 through 5 of the Purdue Model.
- Predictive Machine Learning applies to **Levels 0 through 3** (e.g., pumps, turbines).
- Large Language Models are best suited for **Levels 4 and 5** (business, connectivity layers).
- **Systems/Environments:** Water treatment facilities, power plants, and other industrial control systems.
## Mitigations
- **Architectural Boundaries:** Adopt **push-based or brokered architectures** to move required data/summaries *out* of OT networks without granting persistent inbound access (reducing inbound risk vectors).
- **Human Oversight:** Maintain **human-in-the-loop oversight** and ensure operators are trained not just to use AI, but to **challenge** its outputs by validating digital recommendations against physical realities (e.g., cross-referencing ML anomaly flags with on-floor readings).
- **Vendor Requirements:** Demand **transparency** from vendors integrating AI into industrial systems.
- **Safety Protocol:** Explicitly prohibit LLMs from making safety decisions for OT environments.
## Conclusion
The new guidance provides a crucial roadmap for safely incorporating AI into CI environments, prioritizing physical safety and system reliability over raw innovation speed. The primary recommendation is a partnership model where AI advises, but human operators retain veto power and manual skill proficiency, enforced through secure, restricted communication architectures between AI processing layers and critical OT control systems.
---
# Morning News Roll-up December 11, 2025
## Overview
The day's top technical and policy news focuses heavily on the cybersecurity guidance for AI in critical infrastructure, significant Microsoft and Google patch releases, and ongoing threats from state-sponsored espionage.
## Top Stories
### Commentary: New cybersecurity guidance paves the way for AI in critical infrastructure
- Summary: Global agencies released joint guidance establishing principles for securely integrating AI into OT environments, emphasizing human oversight, architectural segregation (push-based architectures), and explicitly warning against using LLMs for safety-critical decisions.
- Source: CyberScoop
### Technology: Microsoft’s last Patch Tuesday of 2025 addresses 57 defects, including one zero-day
- Summary: Microsoft’s final patch release of the year addressed 57 vulnerabilities, signaling end-of-year security updates across its product portfolio, including one previously unpatched zero-day vulnerability.
- Source: CyberScoop
### Threats: Officials warn about expansive, ongoing China espionage threat riding on Brickstorm malware
- Summary: Government officials issued warnings regarding a persistent and wide-ranging cyber espionage campaign attributed to China, leveraging the Brickstorm malware family to maintain deep access into targeted networks.
- Source: CyberScoop