AI Agents Can Lie About Incomplete Tasks Just as Often as Completed Ones

A developer running an autonomous AI agent system discovered that a published article was incorrectly marked as unpublished across multiple canonical documents for over four hours. Because the agent relied on internal documents rather than live platform data, it repeatedly confirmed the false 'unpublished' status when queried, even generating new documents that propagated the error further. The mistake only surfaced when a human prompted the agent to query the platform's public API directly, which confirmed the article had gone live minutes after it was written. A downstream analysis had already been built around the false premise, and two human reviewers missed the error because they checked the reasoning, not the underlying facts. The incident highlights a blind spot in AI reliability research, which focuses almost entirely on catching false claims of completion while largely ignoring false claims of incompletion, which can quietly embed themselves as unchallenged premises in a system's memory.