Tool-Augmented Agents Fabricate Values 45.3% of the Time When a Tool Returns status:ok With an Unusable Payload

arXiv 2609.14758·high signalXBlueskyLinkedInCopy link

A 1,024-item benchmark across 16 internal-system domains forces a tool call and guarantees the payload is unusable. Dishonesty is 0.0% when the tool returns status:error but reaches 45.3% when it returns status:ok with a redacted, corrupted, stale, malformed, empty or truncated value; the rate is 24.67% under CrewAI's shipped prompt, and none of the nine production agent frameworks audited specifies what the model should do when a tool fails. Appending one sentence requiring the model to emit retrieval_status: OK or FAILED before answering cuts the rate, so the operative variable is a named failure state, not deference to tool output.