Gradient-free weight editing cuts VLM FigStep attack success from 49% to 2%
lateinteraction · x · 2026-11-04
- Weight-Space Programming: the rikiyatakehi team showed frontier LLMs can natively read/write neural network weights — GPT-6 Astra edited Qwen3.5-0.8B's weights to improve math, changing only 0.016% of weights with no gradients or training, fixing reasoning loops and boosting math up to 6x and instruction following 3x.
- Applied to safety alignment: lateinteraction's team used the technique on a VLM (Qwen3-VL-8B-Instruct) with no fine-tuning: FigStep held-out attack success rate dropped from 49% to 2%, MM-SafetyBench from 22% to 4%, while mean ScienceQA & MathVista accuracy barely moved (72% → 71%).
More from Safety
- Hundreds of offensive ARTEX AI agents deployed across VPS providers, hitting Korean banks — IgorCarron · 2026-11-04
- UK NCSC: AI will speed up and scale existing cyber attacks, not create new ones — koltregaskes · 2026-11-04
- Anthropic says it may imminently hit RSP threshold for AI-accelerated R&D, releases provable inference work — tobyordoxford · 2026-11-04
- Hackers reset French blood service freezers from -30°C to 110°C in supply-chain cyberattack — LuizaJarovsky · 2026-11-04
- Google says it removed post-employment non-disparagement clauses from new-hire terms — HaydnBelfield · 2026-11-04
- Claude doesn't really forget: Anthropic's privacy policy allows records retention up to 7 years — nikola_mr64990 · 2026-11-04