Large language models (LLMs) are increasingly used as scientific assistants, but their helpfulness can become harmful when they validate or build upon a user’s flawed scientific premise instead of challenging it. We study this failure mode as scientific sycophancy and introduce SciSycEval, a trap-based benchmark that tests whether LLMs reject, ambiguously qualify, or sycophantically operationalize flawed but plausible premises. Using trap assertions constructed from evidence abstracts across six domains and three reasoning-error types, we evaluate high-performing models under controlled task and authority settings. We find that scientific sycophancy decreases sharply across stronger GPT models, but persists as a trigger-dependent failure mode: generative tasks, high-authority framing, domain-specific evidential gaps, and overgeneralized premises substantially amplify sycophantic behavior. Critical-evaluation prompts reduce but do not eliminate the failures, showing that scientific LLMs must be evaluated and safeguarded for epistemic robustness under the interaction pressures that trigger sycophancy, rather than for factual accuracy alone.