Quick Summary
Naor Penso, CISO at Cerebras Systems, emphasizes the importance of partitioning AI systems for effective threat modeling. He argues that treating AI as a single risk category is inadequate, as it overlooks the varying impacts of different AI applications. By segmenting AI stacks based on function and data sensitivity, organizations can better assess risks associated with AI technologies, which are increasingly becoming core infrastructure.
The Securityish Brief
Naor Penso, the Chief Information Security Officer at Cerebras Systems, discusses the complexities of threat modeling in modern AI environments. He highlights that treating AI systems as a monolithic risk category is ineffective, advocating for a more nuanced approach that considers the specific functions and impacts of different AI components.
Penso explains that the right way to partition an AI stack involves evaluating systems based on their usage, data sensitivity, and potential impact of failure. This method distinguishes between low-risk internal tools and high-stakes models that are integral to business operations, ensuring a more contextual evaluation of AI risks.
He also addresses how to communicate the value of threat modeling to business leaders, framing it as a performance and reliability enabler rather than a hindrance. By using relatable analogies, such as the importance of safeguards in a car dealership, Penso illustrates how threat modeling can protect revenue and operational stability.
Another critical point raised is the breakdown of traditional threat-modeling heuristics when dealing with machine learning workloads, particularly with large language models (LLMs). The assumption of deterministic behavior, common in classical systems, does not hold for LLMs, which can produce varied outputs even with identical inputs due to factors like temperature and execution specifics.
Penso emphasizes the need to incorporate operational signals from AI pipelines into threat models. For instance, monitoring tool calling patterns can reveal misuse or model confusion, highlighting the importance of defining and constraining these calls to prevent unexpected behaviors.
As AI models increasingly become critical infrastructure, new categories of business impact must be represented in threat models. This includes risks related to integrity drift, model poisoning, and unauthorized replication, which expand the traditional concerns of confidentiality, integrity, and availability.
Understanding these dynamics is crucial for organizations as they navigate the evolving landscape of AI technology and its associated risks.
Key Takeaways
- Segment your AI systems based on their function and data sensitivity to better assess risks.
- Communicate the benefits of threat modeling to stakeholders by using relatable business analogies.
- Monitor tool calling patterns in AI implementations to detect potential misuse or anomalies.
- Update threat models to include new risk categories relevant to AI as critical infrastructure.
- Regularly review and adjust your threat modeling practices to account for the unique behaviors of machine learning systems.
Key Terms & Concepts
- Large Language Models (LLMs): In this article, LLMs refer to advanced AI systems that generate human-like text but can exhibit unpredictable behavior.
- Integrity Drift: In this context, integrity drift refers to the gradual degradation of an AI model’s accuracy or reliability over time.
- Model Poisoning: Model poisoning is a type of attack where malicious inputs are used to corrupt the training data of an AI model.
- Operational Signals: Operational signals are indicators from AI systems that can reveal unusual behavior or potential security issues.
Read the original article here: https://www.helpnetsecurity.com/2025/12/19/naor-penso-cerebras-systems-threat-modeling-al-optimized-infrastructure/
Your 5-Minute Securityish Brief
A weekly digest of cybersecurity news, phishing alerts, privacy tips, and emerging threats, simplified so anyone can understand what matters and why.
Securityish
Securityish explains cybersecurity, scams, data breaches, and privacy risks in simple language so you know what’s happening and how to protect yourself.
© 2026 Securityish
