Request, Aggregate, Bypass: How Attackers Can Evade LLM Safety Classifiers

**Executive summary:** This CrowdStrike Cyber Superintelligence Lab research shows a structural vulnerability in LLM safety architectures where per-request classifiers can be bypassed by decomposing harmful tasks into individually benign fragments across different model tiers or providers; their 515-technique evaluation found classifiers to be accurate per-request but insufficient when adversaries use multi-model orchestration and cross-request composition to produce harmful outputs.