Microsoft finds Google and Meta models misled by single prompt

Technology Oct 03, 2026

Microsoft's latest research shows that even top AI models, like Google's Gemma and Meta's Llama, can be thrown off track by a single, relatively mild prompt.

This means safety features we rely on might not always hold up in real-world situations, which is a bit worrying for anyone who uses or builds with AI.

GRPO exposes vulnerabilities, Microsoft urges monitoring

The team found these vulnerabilities using a method called Group Relative Policy Optimization (GRPO), and the issue popped up in text-to-image tools like Stable Diffusion too.

Microsoft says this is a wake-up call: We need to keep checking and updating AI systems after they launch, so they don't slowly start allowing harmful content or risky behavior over time, especially after deployment, when models are built into bigger workflows.