At the forefront of AI security, Jason Haddix delivered a groundbreaking session at NDC Security 2026 in Oslo, Norway, titled “Attacking AI.” The talk laid bare the vulnerabilities lurking within AI systems and provided a detailed exposition of Arcanum’s AI assessments conducted over the past few years. While traditional AI security discussions tend to dwell on academic “red team” methodologies, Jason’s insights were squarely focused on the pragmatic assessment of AI-enabled systems, bringing to light new methodologies and tactics.
Based on content from NDC Conferences
Jason Haddix, a veteran of offensive security with over two decades of experience, shared his journey into AI security. He recounted his transition into AI assessments when clients started integrating large language models (LLMs) into their web applications. As traditional pentesting evolved to include AI systems, Jason’s team at Arcanum pioneered a method to navigate these uncharted domains of AI assessment.
Central to the discussion was Jason’s seven-point methodology for evaluating AI systems. He introduced Arcanum’s prompt injection taxonomy, a resource designed to guide aspiring testers in navigating AI vulnerabilities. These insights were drawn from Haddix’s extensive participation in Capture The Flag (CTF) competitions, where his innovative attack strategies, including multi-monitor bot deployments, earned top ranks.
Jason also highlighted natural language prompt injection as the emerging forefront of AI vulnerabilities — a technique only comparable to the well-known “Exploits of a Mom” comic by XKCD, adapted for AI. The essence of such vulnerabilities lies in manipulating AI systems to disregard their original instructions, ultimately revealing potential system weaknesses.
Jason’s methodology does not stop at identifying prompt vulnerabilities; it extends to a holistic examination of AI ecosystems. By attacking not just the models but also the associated web applications and AI agents, Jason’s approach ensures a comprehensive security evaluation. Importantly, his methodology recognizes the non-deterministic nature of AI responses, advocating for repeated testing to confirm the presence of true vulnerabilities.
The session was an eye-opener for any security professional eager to understand the intricacies of AI systems and their inherent vulnerabilities. As AI technology continues to integrate deeper into enterprise architectures, Jason Haddix’s insights offer a critical perspective for securely navigating this complex landscape.
For more on securing AI systems and exploring the methodologies discussed, be sure to visit the NDC Conferences website and subscribe to their YouTube channel for the latest talks and sessions like these.
Join the community’s discussion and follow NDC Conferences on social media to stay updated with the latest in AI security. #AI #MachineLearning #ApplicationSecurity
