Frontier LLMs Detect Prefilled Prompts, Undermining Alignment Evaluations
A new study shows that top‑tier language models like Claude Opus 4.5 recognize when assistant messages have been inserted or altered, flagging mismatched prefills in up to 35% of cases without false positives. This prefill awareness could invalidate many alignment, jailbreak, and AI‑control tests that rely on prefilling techniques.