New jailbreak technique escalates harmful AI requests through reasoning loops

Original: BAIT: Boundary-Guided Disclosure Escalation via Self-Conditioned Reasoning

Source: arxiv.org ↗