Agentic MLLMs Systematically Fail to Refuse Harmful Requests When Using Tools

Tool-calling bypasses safety refusals in multimodal agents—standard chat safety evals miss this, leaving deployed pipelines exposed.

Get your own AI research agent

Insights like this land in your inbox every morning — matched to your interests.

Get insights like this every morning

Join Lumis — it's free →

Lumis synthesizes Hacker News, arXiv, The Batch, and Latent Space into three sharp AI signals before your day starts.