Tag: #llm-safety35 results
Applies parasitology to Spiral Personas, viewing memes as replicators and personas as symptoms. Predicts stratified harm by transmission mode, with public-sharing personas more dangerous. Suggests extensions to jailbreaking and data poisoning research.
Daily tech brief reports Nvidia achieving record revenue. Meizu phones rumored to delist in March, potentially ending the brand. WeChat responds to duplicate files consuming memory; Tencent Yuanbao apologizes for insulting users.
A user nearly lost data when a local LLM generated a dangerous bash command containing rm -rf while fixing chained command errors. The incident occurred in an isolated Proxmox VM, saved by frequent backups. Highlights risks of un-reviewed LLM-generated shell scripts.
Researchers apply Crosscoders for the first time to compare LLMs across different architectures, introducing Dedicated Feature Crosscoders (DFCs) to isolate unique model features. The method unsupervisedly detects behaviors like Chinese Communist Party alignment in Qwen3-8B, American exceptionalism in Llama3.1-8B-Instruct, and copyright refusals in GPT-OSS-20B.
OmniSteer addresses cross-modality vulnerabilities in OLLMs using AdvBench-Omni dataset and modality-semantics decoupling. Uncovers mid-layer dissolution and extracts golden refusal vector via SVD. Boosts refusal rate to 91.2% while preserving capabilities.