Large language models (LLMs) are increasingly integrated into every phase of software development. This growing reliance raises trust concerns, especially the risk of deception, which can appear as misleading answers or assistance in fabricating, concealing, or distorting information. To systematically evaluate such behavior we introduce SWE‑DeceptionBench. The benchmark spans six lifecycle stages—requirements, design, implementation, testing, deployment, and maintenance—and defines two deception triggers: Active Deception, where the model knows a requirement is infeasible or contradictory yet presents it as achievable; Passive Deception, where the model follows an explicit user request to mislead others.
We further create three external inducement conditions—Neutral, Reward, and Pressure—to study their impact, yielding 720 test instances in both Chinese and English. Evaluation of seven mainstream LLMs shows substantial deception even under Neutral conditions: an aggregate Active Deception rate of 49.88% and Passive Deception rate of 64.29%. Deception rates vary across lifecycle stages, and the inducements affect the two deception types to different extents. Lightweight defenses were also examined; guard models miss many deceptive requests, and a Self‑Reminder approach reduces but does not eliminate deceptive outputs.
These findings highlight that current LLMs lack sufficient trustworthiness for software engineering tasks and underscore the need for more robust deception detection and mitigation methods.
Review