Rebuilt weight-space safety edit cuts Qwen3-VL-8B FigStep attack success from 28.2% to 0.2%
kevrussell · x · 2026-11-05
- @honestytools rebuilt a weight-space safety edit on Qwen3-VL-8B following the original authors' published log.
- FigStep jailbreak attack success dropped from 28.2% to 0.2%, while ScienceQA and MathVista scores stayed flat.
- The cost: the edited model now refuses over half of harmless image-text requests, showing a severe over-refusal trade-off in weight-level alignment edits.
More from Safety
- HF dev questions Thinking Machines' SWE-2 Flash weight-sharing program over simple HF gating — QGallouedec · 2026-11-05
- OpenAI explicitly endorses stronger chip and SME export controls — ShakeelHashim · 2026-11-05
- What port forwarding actually is — and why agentic scanners make it riskier — victorialslocum · 2026-11-05
- AI Is Taking Over Corporate Decisions — Humans May Keep Authority but Lose Control — uxmag · 2026-11-05
- a16z: AI Cut Exploit Writing to Hours as Attacks Now Land 7 Days Before Disclosure — a16z · 2026-11-05
- EU AI Act Scientific Panel names new Chair, brace for 18 months of systemic-risk work — S_OhEigeartaigh · 2026-11-05