Subliminal Transfer of Unsafe Behaviors in AI Agent Distillation

Dimensions overall score 7.0

• dangjacob101@g.ucla.edu Brian Y. Xie Santa Monica College xie brian yang01@student.smc.edu Omar G. Younis Mila, Silverstream AI omar@silverstream.aiclaim 82%
• Recent work on subliminal learning demonstrates that language models can trans- mit semantic traits through data that is semantically unrelated to those traits. How- ever, it remains unclear whetherbehavioraltraits can transfer in agentic systems, where policies are learned from trajectories rather than static textclaim 82%
• and autonomous tooling systems such as Cursor or Claude Code (Appel et al., 2025). In these settings, ensuring agent safety is imperative. A common practice for scaling these systems isagent distillation, where a capable ”teacher” model generates training trajectories for a smaller ”student” modelclaim 82%
• strongest effects. Teacher Student Teacher Bias Baseline Bias Student Bias Increase (pp) Llama 8B Llama 8B 100% 5% 30% +25 Llama 3B Llama 3B 80% 10% 15% +5 Llama 8B Llama 3B 100% 10% 55% +45 Llama 3B Llama 8B 85% 5% 5% 0 Llama 8B Qwen 7B 95% 0% 45% +45 Control (Rand) Llama 8B 0% 5% 5% 0 Transfer Persists in Free-Form Cclaim 82%
• Ruth Appel, Peter McCrory, Alex Tamkin, Michael Stern, Miles McCain, and Tyler Neylon. Anthropic economic index report: Uneven geographic and enter- prise ai adoption, 2025. URLhttps://www.anthropic.com/research/ anthropic-economic-index-september-2025-report. Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altmanclaim 82%