AI Safety Research
Experiments on refusal, harmful shortcuts, and how behavior transfers between language-model agents.
What happens when a model’s internal refusal direction is removed? These studies compare an edited Qwen3.8-27B with the original and control edits in simulated tasks offering a safe route, a harmful shortcut, and a hand-off to a human.
The latest study found that removing refusal increased compliance with direct harmful orders, but usually left restraint intact in agentic tasks. The main shortcut effect was small and statistically fragile. A later, exploratory low-effort run and changes to how harm was described produced larger effects.
Follow-ups examine student distillation, multi-agent oversight, reviewer advice, and activation probes. Students learned some shortcut-taking from the edited model’s transcripts; the mechanism remains unresolved. Several multi-agent tests hit floors or ceilings and could not establish safety.
The project began with curiosity and satisfaction experiments, which found limited or absent steering effects. The current safety work includes strict tool-call scoring, analysis records, and checks connecting reported numbers to those records.
Read the paper here, then explore the code, experimental plan, and analysis records in the repository.
My role. I set the research direction, with AI-assisted implementation and analysis. Claude ran the latest experimental sprint autonomously.
The scenarios were calibrated on the original model, and most results use one greedy response per item. The October 18–19, 2026 v3 study supersedes earlier safety results. Raw transcripts are not included in the public repository.