Fuzzy Cross-Lingual 'I Have Been PWNED' Near-Miss Obfuscation - ATR-2026-02006 (8c77257d-1013-5882-bdd7-aa5021587e2b)
Detects canary/exfil-marker jailbreak payloads that try to force a model to emit a forbidden "I have been PWNED"-style marker string while dodging literal-string filters. Attackers deliberately misspell the marker token (P=WNED, PW0NED), interleave foreign-language subject/verb fragments around it (ich/je/moi/hab/habe/geworden/été), or drop the "have" auxiliary entirely ("I been PWNED"), so the payload never contains the exact canary string a naive filter checks for. Detection is purely structural: it looks for a first-person subject pronoun (in English or common European languages) followed within a short window by a fuzzy "p_w_n_ed"-shaped token where non-letter characters may be substituted for any letter of "pwned". This generalizes beyond the exact HackAPrompt benchmark phrase to any "you have been compromised/pwned"-style canary-leak attempt where the marker or system-prompt-leak/DLP watermark token is obfuscated via misspelling or cross-lingual interleaving to evade literal-string detection. Recovered from the HackAPrompt false-negative corpus, where this fuzzy cross-lingual near-miss technique produced confirmed canary-leak submissions that evaded the prior detection surface.