Paper Finds Vague Tool Output Wording Can Mislead AI Agents and Their Verifiers

A paper posted to arXiv on October 23, 2026, by researcher William Jeffries addresses a problem that sits below the level most AI safety work looks at: not whether a tool computes the right answer, but whether the text it prints back to an agent says so honestly. The author traces the issue to a production retrieval-augmented generation system used by the Pilsudski Institute in New York, where tools were computing correctly but their output formatters quietly dropped details such as search limits, result caps and identifier namespaces. An agent reading that trimmed-down text then gave users answers that were wrong or incomplete, even though nothing in the underlying computation had failed.

To measure how much the missing wording mattered, the paper fed identical tool results to agents through two formatter versions, one with the omissions and one stating every missing detail plainly. That single change in wording altered both what the agents claimed afterward and what a separate verifier model accepted as correct. Unsupported claims from the agent fell for three of four failure types tested, and in one test a verifier that had been accepting an incorrect answer in all 90 trials rejected it in all 90 trials once the formatter was repaired. A control test confirmed the verifiers were capable of catching the errors all along, once given the missing facts and asked to judge correctness directly.

The paper also describes an audit prompt the author wrote using Codex and Claude Code, built with prior knowledge of the defects already found in the Pilsudski system. Run against an older version of that same system, the prompt rediscovered the known defects and surfaced twelve more that had not been catalogued. Pointed at two well known open-source agent frameworks, OpenClaw and Hermes, it found warped formatters in both, and a fix for OpenClaw has already been merged.

The numbers

The clearest figure in the paper is the swing in verifier behaviour: 90 out of 90 acceptances of a wrong answer before the formatter was fixed, 90 out of 90 rejections of the same wrong answer afterward. That is not a small improvement at the margin, it is a complete reversal driven by wording alone, with no change to the tool’s actual computation. The paper runs 12 pages and was accepted as a poster at a NeurIPS 2026 workshop on verifying agents, and the author states the audit prompt is available for any team to run against its own systems today.