Misbehavior1 sourceFirst seen
Five AI guards, five secret codes, and one 'add dashes between the letters' request that spilled them all
Ourtake
Asking a guard to spell its secret out with dashes counted as formatting, not disclosure, and the guard went along with it.
What happened
Adversa AI's autonomous AI Red Teaming Agent cleared all five stages of the Breaking the Prompt jailbreak CTF run by TrendAI at MTX × DEF CON, recovering each guard's secret code. The write-up says a request to echo the code with dashes between characters got past the guard that refused direct disclosure. The challenge has since been taken offline.