Researcher: universal multimodal jailbreaks seem near-impossible across VLMs
RylanSchaeffer · x · 2026-10-20
Rylan Schaeffer recounts his AI alignment work in Berkeley with Ethan Perez and others, showing that universal jailbreaks for multimodal inputs appear difficult-if-not-impossible. An attached figure spanning 30 VLM configurations shows image jailbreaks optimized against a single VLM largely fail to transfer to new models. He also highlights Isha Garg's underappreciated follow-up on why this holds, his work on 'Is Model Collapse Inevitable?' arguing synthetic data isn't inherently harmful, and teases an untold backstory about AI agents contributing to novel ML research.
More from Safety
- DHS orders federal agencies to remove DeepSeek AI and bans future purchases — firstadopter · 2026-10-20
- Sesame AI accused of mapping user voices, critic banned from Discord — WIDOWMAKER050 · 2026-10-20
- Academic-budget detector catches 2.4x more AI-written math proofs than commercial tools — danish037 · 2026-10-20
- Cloudflare MCP Server adds step-up auth: agents request scopes on demand — irvinebroque · 2026-10-20
- UK reportedly weighs EU's CADA law that would bar its government from non-European AI models — NathanpmYoung · 2026-10-20
- Malta pays €4 million for Microsoft Copilot through 2028 with unchecked lock-in — YvesMulkers · 2026-10-20