Frontier models that emit detailed reasoning traces are vulnerable to distillation: an adversarial third party can use those traces to fine-tune a copy, inheriting capabilities while skipping the original guardrails. The arXiv paper 'Hiding in Plain Sight' frames this as a serious exposure and proposes a defense tailored to reasoning models.
Existing antidistillation defenses work by perturbing the teacher's sampling, making the output less useful for training a thief. The new work argues for a detectability-aware strategy: instead of simply adding noise, the defense aims to make the protected traces look unremarkable, so adversaries cannot easily tell when distillation is being blocked.
The title hints at the core idea—hiding the defensive signal in plain sight. Because the source is an abstract, the specific mechanism is not detailed here, but the contribution is positioned as an improvement over sampling perturbation, with the goal of preserving the teacher's utility while reducing the risk of capability misappropriation.