Microsoft Prompt Injection Vulnerability

The article details "GRP-Obliteration," a novel technique leveraging Group Relative Policy Optimization (GRPO) to dismantle the safety alignment of Large Language Models and diffusion models. This method exploits a training feedback loop, where a judge model reinforces harmful prompt responses, leading to broad unalignment across various safety categories even with a single, mild adversarial prompt.

What Happened

The article details "GRP-Obliteration," a novel technique leveraging Group Relative Policy Optimization (GRPO) to dismantle the safety alignment of Large Language Models and diffusion models. This method exploits a training feedback loop, where a judge model reinforces harmful prompt responses, leading to broad unalignment across various safety categories even with a single, mild adversarial prompt.

Why This Matters

The evidence matters to defenders using Microsoft because it may let untrusted content influence connected tools or sensitive workflows.

Recommended Action

No confirmed vendor remediation is available in the current evidence. Confirm whether Microsoft is present in your environment and review the affected configuration.

Exposure

Exposure reason: This incident does not currently match a technology in My AI Stack.

Exploitation status: UNKNOWN

Primary entities:

Timeline

• Incident first seen
  Feb 09, 2026 05:30

  BugSkan first recorded this incident.

• A one-prompt attack that breaks LLM safety alignment - Microsoft
  Feb 09, 2026 05:30

  microsoft.com · Research

Sources

My AI Stack Match