AI Red Teaming: Overcoming the Willing Attacker Problem

A10 Networks

Nov 10, 2026

In this AI red teaming session, Jamison Utter and Priyanka Mullan from A10 Networks explain how to solve the willing attacker problem when testing generative AI defenses. Because safety-aligned large language models often refuse adversarial prompts or introduce unintended hallucinations, security teams use a process called model obliteration to suppress refusal behavior and remove built-in safety guardrails from testing models. This enables realistic attack iteration while keeping production target defenses fully intact.

Key Takeaways:

Safety-aligned models refuse attack-exploration prompts due to built-in harm avoidance.
Extraneous model inference often triggers hallucinations during complex scenario generation.
Model obliteration removes refusal code and basic safety guardrails to enable effective adversarial simulation.
A stronger simulated attack is not automatically a finding until teams verify what the production target actually disclosed or executed.

Learn more about AI Red Teaming: https://bit.ly/4srUI1W

Timestamps

00:00 The Willing Attacker Problem in LLM Testing

00:09 How Refusal Mechanisms and Inference Cause Hallucinations 00:27 Using Obliteration to Strip Constraints for Adversarial Attacks

00:54 A10 Networks Overview

#airedteaming #aisecurity #a10networks #cybersecurity #llmsecurity #adversarialai #obliteration #machinelearning