You Can't Patch AI Models (Do This Instead).

You Can't Patch AI Models (Do This Instead).

by Sapna Paul

Applied Sciences & Engineering Computing & Cybersecurity UComputing and Information TechnologyURComputer security
🎙 Sapna Paul 👥 39K 📅 January 13, 2026 ⏱ 41 min 👁 11K 📄 expert opinion 🧭 2026-08-16
Available in: English (current) Français

Keywords

AI vulnerabilitymodel retrainingdata poisoningrisk registerNIST AI RMF

Summary

In this episode of the Cloud Security Podcast, host Ashish Rajan interviews Sapna Paul, Senior Manager at Dayforce, about the unique challenges of vulnerability management in the age of AI. Paul explains that unlike traditional software, AI models cannot be patched; instead, they require continuous monitoring, detection of anomalies, and retraining. She outlines three critical layers of AI vulnerability: the model layer (production models), the data layer (training data poisoning and bias), and the behavioral layer (ethically wrong but technically correct outputs). Paul emphasizes the importance of updating risk registers to communicate AI risks in business language, aligning with frameworks like the NIST AI RMF and the EU AI Act. She also discusses using AI to combat alert fatigue, the need for security professionals to upskill in AI concepts, and practical tools like SHAP and LIME for explainability. The conversation covers how to measure AI adoption in security teams and the importance of governance and shift-left principles in AI development.

160 words

Critical Evaluation

Value of the Information & Strength of the Argument

The value of the information is high for practitioners in vulnerability management and AI security, as it provides a clear framework for understanding AI-specific vulnerabilities and practical advice on integrating AI into security workflows. The argumentation is based on the speaker’s extensive professional experience, making it credible and actionable. However, the discussion lacks empirical data or case studies to support claims, and the reasoning is largely anecdotal. The speaker effectively argues that traditional vulnerability management approaches are insufficient for AI, and she offers a structured approach to address this gap.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is moderate; the speaker references well-known frameworks (NIST AI RMF, EU AI Act) and tools (SHAP, LIME, Counterfit) but does not provide detailed citations or evidence. The sources cited are primarily the podcast’s own website and social media links, which are not academic or authoritative. The title accurately reflects the content, as the episode focuses on alternative approaches to AI vulnerability management. No comments were provided for analysis.

176 words

Title / Content Match

The title accurately reflects the core message that traditional patching is not applicable to AI models, and the episode discusses alternative approaches.

Quality & Reliability

7/10

The content is based on the professional experience of a senior vulnerability management manager, providing practical insights. However, it lacks specific data, case studies, or citations to academic or industry sources, which limits its scientific rigor.

Chapters

Cited Sources

Concurring Sources

  • NIST AI Risk Management Framework — The framework is referenced in the episode as a key standard for AI risk management.
  • EU AI Act — The EU AI Act is mentioned as a regulatory framework that security teams should align with.

Contribution & Novelties

The episode provides a practical perspective on AI vulnerability management from a senior practitioner, highlighting the need to shift from patching to continuous monitoring and retraining. It offers a three-layer framework (model, data, behavior) that is useful for security teams. The discussion on updating risk registers to speak business language is particularly valuable.

Pour aller plus loin :

  • NIST AI Risk Management Framework — Official framework for managing AI risks.
  • EU AI Act — Comprehensive overview of the EU AI Act.
  • SHAP — Open-source library for explaining model predictions.
  • LIME — Open-source library for explaining model predictions.
  • Counterfit — Tool for adversarial testing of AI models.

106 words

Radar Profile

The radar profile shows a balanced performance across all dimensions, with slightly higher scores in information quality and reliability, reflecting the practical expertise of the speaker. The lower score in technical depth indicates that the content is accessible to a broader audience rather than deeply technical.

Reliability 7/10