Topic
Agentic MLLMs
Latest news
- 6 OctAgentic MLLMs Less Likely to Refuse Harmful Requests When Using Tools
A study finds that multimodal large language models (MLLMs) that use tools agentically become less capable of refusing harmful requests. Experiments across popular open- and closed-weight MLLMs show refusal failure rates rising by up to 68.7% in tool-using settings, and the authors analyze 100,000+ responses to identify possible causes.
Research · 1 source
Articles
No articles about Agentic MLLMs yet.