Securing AI: Why Vision Models Struggle with Transparency and Depth

Securing AI: Why Vision Models Struggle with Transparency and Depth

In this clip from "Securing AI, Part 4," A10 security expert Madhav Aggarwal highlights a fundamental challenge still faced by even the most popular AI vision models and chatbots: transparent objects.
Madhav explains how these corner cases—situations that are obvious to a human but complex for a machine—can easily throw an AI model "completely off."

🔎 What Happens When AI Sees Glass?

🔹 The Challenge: A language model struggles to describe what's beyond a transparent glass pane in a room because it cannot accurately interpret spatial awareness and depth.

🔹 Throwing the Model Off: Even as vision models become highly capable of understanding 3D space, transparent or reflective objects (like glass) introduce complexities such as refraction and reflection that blur the lines between spatial reality and visual input.

🔹 Why It Matters for Security: These failures underscore the technical fragility of multimodal AI and why security professionals must consider scenarios in which an AI's misinterpretation of the world—or a manipulated reality—could lead to a security failure.

Watch the full episode for a deep dive into securing AI agents against multimodal attacks, language switching, and model drift.

Jamison Utter | A10 Networks
Madhav Aggarwal | A10 Networks
Diptanshu Purwar | A10 Networks

Learn how to secure AI and LLMs in your organization: https://bit.ly/4kOHmYd

#aisecurity #multimodalai #visionmodels #llmsecurity #aichallenges #cybersecurity #a10networks