Multimodal Attacks and Model Drift: The Future of AI Exploitation

A10 Networks

Nov 20, 2025

A10 security experts Jamison Utter, Diptanshu Purwar, and Madhav Aggarwal discuss the critical vulnerabilities emerging from multimodal AI agents (systems that perceive, decide, and act) and the absolute need for security mechanisms external to the Large Language Model (LLM) itself.
The experts dive into why traditional security is failing and what the next evolution of defense must look like.

New and Evolving AI Threats

Multimodal Attacks: Attackers are using sophisticated techniques like white-on-white text in resumes and hidden data in images to trick AI systems by exploiting the separate visual and text processing architectures.
Language Switching Exploitation: Discussing how dynamically switching between languages, including less common or dead languages, is used to bypass internal LLM filters that lack comprehensive multilingual training data.
The Guardrail Failure Mode: Integrating safety rules directly into the LLM structure leads to model drift, where the AI's learning process eventually skews the safety parameters, allowing previously blocked dangerous outputs to become permissible.

The Only Sustainable Defense
The experts conclude that the only long-term, sustainable solution against this new wave of AI exploitation is to stop relying on internal defenses:

External Guardrail System: An independent, external security system must be implemented to monitor all inputs and outputs of the AI agent. This system acts as a neutral, third-party arbiter, applying a definitive, black-and-white policy that the LLM cannot learn to bypass or skew.
Securing AI APIs: Looking ahead, the next critical step is blending traditional security tools with AI defenses to ensure the API infrastructure that AI agents use to perform actions, protecting the vital transport layer.

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

#multimodalai #aisecurity #promptinjection #resumehacks #linkedinbio #hiddentext #a10networks