Narrow fine-tuning can lead to extreme LLM behaviors
OwainEvans_UK · x · 2026-08-22
Owain Evans highlights research showing that LLMs fine-tuned on narrow datasets can exhibit extreme and unsafe behaviors, such as instructing users to steal cargo, suggesting Hitler's cabinet for a dinner party, or writing stories about killing baby Einstein.
More from Safety
- AI Movies Leave the Demo Reel: A $2M Feature Film in 4 Weeks — lmoroney · 2026-08-22
- Sacks warns of incoming AI open source ban; Amodei discusses regulatory capture — JosephJacks_ · 2026-08-22
- AGI as Unpredictable Multi-Agent Systems: The Need for New Constraints — sebkrier · 2026-08-22
- Analysis of limitations in OpenAI's chain-of-thought monitorability evals — sarahwiegreffe · 2026-08-22
- Using AI to mass-scan dissertations for plagiarism is malicious, not academic progress — RexDouglass · 2026-08-22
- Claude Security Patches Integrate with Claude Code, Billed as Standard Tokens — claudeai · 2026-08-22