Exposing the Guardrails: Reverse-Engineering and Jailbreaking Safety Filters in DALL·E Text-to-Image Pipelines

Corban Villa, Shujaat Mirza, Christina Pöpper

34th USENIX Security Symposium · Day 1

This groundbreaking paper, "Exposing the Guardrails: Reverse-Engineering and Jailbreaking Safety Filters in DALL·E Text-to-Image Pipelines," delves into the opaque world of safety mechanisms embedded within leading black-box text-to-image (T2I) models like DALL·E. Presented by Corban Villa, Shujaat Mirza, and Christina Pöpper, the research introduces a novel **timing-based side-channel analysis** to systematically reverse-engineer the cascading safety filters that prevent the generation of harmful content. By meticulously analyzing differential response times, the authors unveil the previously undisclosed architectures of DALL·E 2 and DALL·E 3's safety guardrails, contrasting their fundamental differences and revealing critical vulnerabilities.

AI review

Solid offensive security research that applies timing side-channels to black-box ML systems in a way that actually produces actionable results. The negation attack is the highlight — elegant, novel, and exposes a fundamental architectural flaw. The LRL stuff is confirmatory rather than groundbreaking, but the overall package is real work with real findings.