Request, Aggregate, Bypass: How Attackers Can Evade LLM Safety Classifiers
CrowdStrike, Tuesday, October 6th, 2026
CrowdStrike researchers bypassed a leading LLM safety classifier by splitting harmful requests into benign subtasks.
The CrowdStrike Cyber Superintelligence Lab evaluated the most advanced publicly deployed content safety classifier guarding frontier models and found it highly robust against direct attacks.
However, the classifier could be systematically circumvented by decomposing harmful requests into benign subtasks and reassembling the results, a technique validated across 9 of 10 offensive security categories.
CrowdStrike notes that Microsoft Research independently published a similar "Capability Laundering" attack in September 2026 and says per-exchange filtering has structural limits.