New defense system stops AI jailbreak attacks without blocking helpful requests

Original: NeuroArmor: Safe-Variant-Guided Representation Consistency for Selective Re-Anchoring in Jailbreak Defense

Source: arxiv.org ↗