Controllability-Aware Adversarial Examples Against LLM-Based Network Traffic Classifiers

A set of arXiv papers (July 10, 2026) that report new attacks, measurements, datasets, and defenses across LLM and related ML-enabled systems. Key findings: LLM-based network-traffic classifiers are substantially vulnerable to controllability-aware black-box transfer attacks (500k+ examples, multiple datasets and targets); text-to-image models in the wild include high-risk derivatives and detector metrics overestimate real-world risk; LLM jailbreaks can be diagnosed mechanistically via paired internal computation graphs enabling targeted interventions; code-completion models remain at risk of/