Data Poisoning

1 article on Data Poisoning.

Subscribe by RSS

• ProductSecurity

  What GRP-Obliteration Tells You About Your AI System's Threat Model

  February 16, 2026

  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.