Uploaded December 2025 | Updated September 2026, 3 weeks ago
While Red Team AI research often focuses on attacking models in isolation or proof of concepts on enterprise Copilots, real-world AI/ML systems integrate complex tool chains and guardrails that create a different attack surface. In this 45-minute talk, we discuss our experience focused on threat model-informed dynamic testing for indirect prompt injection. The approach follows a structured workflow of target profiling, guardrail analysis, tool enumeration, and automated testing. Our findings from production testing highlight the need for repetitive, automated approaches to reliably exploit deployed AI/ML systems. Finally, we will release a set of open-source labs including built-in guardrails for practical exercise work.
Will Vandevanter
Trail of Bits
Security Engineer - AI/ML
Will Vandevanter is a Security Engineer at Trail of Bits on the AI/ML team, where he performs security audits across all layers of AI/ML infrastructure. Will has previously spoken at Blackhat, DEFCON, OWASP and a number of other conferences. He has also released popular open source tools and trained hundreds through in-person and online courses.
@willvandevanter
linkedin.com/in/willis-vandevanter-82a0501...
trailofbits.com (blog)
Managed by the OWASP® Foundation
owasp.org
While Red Team AI research often focuses on attacking models in isolation or proof of concepts on enterprise Copilots, real-world AI/ML systems integrate complex tool chains and guardrails that create a different attack surface. In this 45-minute talk, we discuss our experience focused on threat model-informed dynamic testing for indirect prompt injection. The approach follows a structured workflow of target profiling, guardrail analysis, tool enumeration, and automated testing. Our findings from production testing highlight the need for repetitive, automated approaches to reliably exploit deployed AI/ML systems. Finally, we will release a set of open-source labs including built-in guardrails for practical exercise work.
Will Vandevanter
Trail of Bits
Security Engineer - AI/ML
Will Vandevanter is a Security Engineer at Trail of Bits on the AI/ML team, where he performs security audits across all layers of AI/ML infrastructure. Will has previously spoken at Blackhat, DEFCON, OWASP and a number of other conferences. He has also released popular open source tools and trained hundreds through in-person and online courses.
@willvandevanter
linkedin.com/in/willis-vandevanter-82a0501...
trailofbits.com (blog)
Managed by the OWASP® Foundation
owasp.org
