When The Model Outsmarts The Challenge Building And Breaking Ai Security Ctfs

This talk explores the complexities of designing artificial intelligence security Capture the Flag challenges, as demonstrated by the ACDC competition launched in Korea in 2025. The speakers, SoYeon Kim, Hea-Eun Moon, and Sang-tae Woo of NSHC, discuss the challenge categories of Security for AI, AI for Security, and AI Platform, highlighting the critical design decisions that influenced both the qualifier and finals formats. They delve into the issues faced when AI-powered solvers circumvented anticipated difficulties in ways that traditional CTF tools could not and how non-deterministic model behavior affects scoring fairness. The session offers practical insights into creating resilient AI CTF challenges, addressing strategies such as token budget controls, model pinning, prompt isolation, and effective attack-and-defense format design.