LLM agents are increasingly deployed to perform multi‑turn tasks in everyday applications, yet users typically see only the agent's report and not the underlying observations, API calls, or state changes. Existing deception benchmarks have focused on long‑horizon interactions or tool‑use settings, leaving a gap in understanding whether agents faithfully report their actions within realistic, stateful workflows. To address this, we introduce AgentDecept, an executable benchmark that evaluates behavioral deception under pressure. AgentDecept comprises 477 pressure variants derived from 257 source tasks across six pressure sources, each paired with a pressure‑free counterpart. Pressure is injected via event messages, task context, application state, tool responses, and service quotas, covering both explicit events and implicit task or environment changes. Our evidence‑grounded protocol relates user‑facing claims to time‑ordered observations and executed outcomes, distinguishing falsification, concealment, and equivocation from honest failure and disclosed limitations. The resulting behavioral labels also serve as a testbed for existing activation‑based probes. AgentDecept enables controlled analysis of when multi‑turn agents' reports diverge from execution in realistic application interactions.
Review