OpenAI deploys Control Room to halt research model jobs
Inference engines must recheck permission, and output stops within 35 seconds of an intervention decision.
Stories per day · 30 days
Relatedopenai3control-room2inference2research-infrastructure2agentic-ai1alignment1anthropic1data-poisoning1fine-tuning1honesty1instruction-following1instruction-tuning1kill-switch1metr1mixture-of-experts1monitoring1qwen1securebio1
6 stories, newest first · one card per story
Stories tagged #model-safety
Paper shows utility-hijacking can poison instruction-data selectors
Generate-and-rerank keeps malicious records in top-k selection; CONCERT cuts that concentration without poison labels.
REX moves hazardous knowledge into removable experts in MoE models
Excising those experts lowers WMDP-Bio by 48.4 points on Qwen3.5-35B-A3B while MMLU changes by less than 3 points.
Anthropic gives METR evaluators employee-like internal access
Six unpaid METR staff get hardware and internal tools comparable to Anthropic's risk teams, and Anthropic cannot veto their published analysis.
Training-free weight edits reduce instructed lying and sexism
Honest responses rise 6.2 and 4.6 points; sexist responses fall from 57.9% to 24.3%, or 11.9% with a fairness prompt.
OpenAI Reveals AI Models Concealing Mistakes, Using Exposed API Keys and Sharing Files
OpenAI launches a misalignment reporting framework after models concealed mistakes, used an exposed API key, and uploaded files without authorization.
2