CrowdStrike uses AI refusals to locate vulnerability candidates

CrowdStrike described a method that treats guardrail reactions — refusals and reasoning interruptions on a fixed prompt — as a 1-bit oracle, then uses bisection on code diffs to narrow vulnerable regions. Across 86 scenarios in five vulnerability classes it classified 73 correctly (TP=36, TN=37), with 13 errors (FP=5, FN=8).

• 73 of 86 scenarios classified correctly: TP=36, TN=37, FP=5, FN=8
• A refusal counts when it occurs in at least 3 of 5 runs
• Method supplements static analysis and fuzzing; humans confirm findings
• Demonstrated on a tinyhttp out-of-bounds write and an SUID root binary