A new arXiv paper highlights a subtle failure mode for AI agents that make decisions in a single forward pass. Such models combine live environmental observations with the agent's own record of previous actions. Normally these sources agree, but when an action fails silently, they can point in different directions.
The authors call the resulting problem "claim laundering": the model appears to act on reality, but is actually acting on its own internal record, which may no longer match the environment. Because the decision is made in one pass, there is no opportunity to notice the discrepancy before acting.
The paper is described in the abstract as introducing this term and analyzing the conflict between records and reality. The full details of the proposed mechanisms or mitigations are not available in the abstract, but the core observation is clear: agents that trust their own history over current evidence can make confidently wrong choices.