What GRP-Obliteration Tells You About Your AI System’s Threat Model
Microsoft researchers reversed LLM safety alignment across 15 models using a single adversarial training prompt. The attack exploits a fundamental design flaw: safety implemented as learned behavior can be inverted through training-based attacks. For product security teams deploying AI systems, the lesson is architectural – security properties must be enforced through system design, not statistical tendencies.