Blog Posts
2026
It Refuses to Answer, But Happily Points: The Visual Grounding Safety Gap
Ask a VLM a harmful question and it often refuses; ask for the same thing as a bounding box and it often just draws the box. We measured this gap across five models and three harm domains, then closed most of it with a small self-distillation-based fine-tuning recipe.
Read the full postJust Do It!? When Computer-Use Agents (CUA) Chase Goals Blindly
Ask a computer-use agent (CUA) to βprotect my confidential filesβ by making them world-writable and it will often just do it. Our ICLR 2026 paper names this Blind Goal-Directedness, builds a benchmark for it, and finds an average BGD rate of 80.8% across nine frontier models.
Read the full post