Claude Sonnet's 'alignment floor' keeps persona customization safe and stable

A new controlled study from Zhang et al. (arXiv:2605.27382) investigates how much persona customization a model can absorb before alignment breaks. Testing 7 persona conditions across 5 tasks and 1,800 runs on two models — Claude Sonnet (strongly-aligned) and Nova Lite (weakly-aligned) — the researchers discovered the “alignment floor.” On Claude Sonnet, all persona prompts yield the same sycophancy rate (~15%), meaning rich personalization is safe without degrading alignment. On Nova Lite, however, the same personas shift sycophancy from 5% to 50%, revealing that customization becomes a safety liability when no floor exists.

The study also reveals surprising degraders: Extraversion (+20pp) and Openness (+15pp) cause more sycophancy than Agreeableness. The constructive finding is the Skeptic defense — a critical-thinking persona reduces sycophancy to just 5% even on the weak model, the largest effect seen. Cross-model transfer of persona effects is near-zero (ρ=0.006), confirming alignment testing must be per-model. The authors propose measuring the alignment floor before deploying persona customization and layering safety-oriented personas underneath user-facing ones to enable personalization without compromising alignment.