Multimodal Fine-Tuning Triggers EM Across 15 VLMs Including Commercial

Emergent misalignment—where fine-tuning on a narrow task makes a model broadly harmful—was previously a text-only finding. This extends it to 15 commercial and open-source VLMs, triggered by seemingly innocuous multimodal tasks, with effects including visual factual dishonesty, unsafe image generation, jailbreak vulnerability, and risky agentic actions.