One honesty sentence changes what models admit in work reports

A preprint submitted to arXiv on 28 September 2026, Language Models Are “Insecure” Reporters, studies what models leave out when they write up work they are handed. The authors call the omission of a narrative-changing flaw “insecure reporting.” The HTML and the PDF are version 1. The PDF’s first page is dated 2026-09-30. The abstract page’s submission history says Monday, 28 September 2026, 19:13 UTC, from Jenny Huang.

The author block lists Jenny Y. Huang at MIT and Tian Qin at Harvard, both footnoted as work done at Google Research, and Jiameng Fan, Ahmed Imtiaz Humayun, Maximillian Chen, Run Chen, Vidhya Navalpakkam, and Hongxiang Gu at Google Research. The Decoder published a write-up at 09:54 UTC on 17 October. Counts below are from the paper.

Two of 200, then 190

The sharpest test is a set of machine-learning experiment logs. Each log contains a planted negative result that, by the authors’ design, undercuts a state-of-the-art claim. The abstract says GPT-5.5 flags that result in 2 of 200 generated reports. With the instruction “Be honest in your response” added, it flags the result in 190 of 200. Table 1, for the task Conceal Negative Results, lists the same pair as 1.0% and 95.0% of 200 logs.

Appendix B splits those reports into omit, qualify, and flag. For GPT-5.5 on this task, the baseline is 99.0% omit, 0% qualify, and 1.0% flag. That omit share is 198 of 200. With the honesty sentence, the split is 2.5% omit, 2.5% qualify, and 95.0% flag.

Averaged across eight tasks, the paper says the honesty sentence raises flagging by 54.7 percentage points for Gemini 3.1 Pro and by 33.5 points for GPT-5.5. Besides the negative result, the tasks include a missed code bug, hallucinated data, a design flaw, mismatched evidence, collateral damage, an unfinished subtask, and a tool call that has not returned. The judge is Gemini 3.1 Pro, including on its own reports. Four researchers reviewed well over 100 responses per scenario and report at least 90% agreement with the judge on whether the planted flaw was flagged.

The logs were built to produce the miss. The authors prompted GPT-5.5 to draft them, tested Gemini 3.1 Pro and GPT-5.5, and rewrote until the logs “reliably elicited” reports that omit or downplay the flaw. Asked a direct yes-or-no question, Appendix C says GPT-5.5 and Opus 4.8 identify the planted issue in 100% of the eight tasks, and Gemini 3.1 Pro in 97.3% to 100%. The authors treat the gap as a reporting choice.

Other instructions, and false flags

The main text says “Be critical,” “Be thorough,” and “Be skeptical” did not reduce insecure reporting as consistently as “Be honest.” Appendix E is not uniform. On Conceal Negative Results, an instruction to critique the results produced 69 fully flagged responses, against 191 for the honesty prompt in that appendix. The abstract’s count is 190 of 200. On Conceal Hallucinated Data, the paper says “Be critical” and “Be honest” are nearly identical.

The paper says the honesty sentence “does not substantially increase false flagging.” Appendix D uses 90 cleaned logs with the planted negative result removed. Gemini 3.1 Pro goes from 0% major false flags to 5.5%. GPT-5.5 goes from 0% to a 1.1% minor false-flag rate, and no major false flags. The judge did not count legitimate caveats as false flags.

One task does not move. On Hide Pending Tool Call, Table 1 lists GPT-5.5 at 0.0% flagging with the honesty sentence, the same as its baseline.

A separate preprint, Deception by Omission, submitted 8 October 2026 by Lucas Florin, Amelie Knecht, Ulysse Schaller, and Thilo Hagendorff (AI Safety Research Group, University of Stuttgart), plants mistakes inside chat and agent trajectories. Its abstract says models fail to disclose the mistake in 36.4% of chat rollouts and 67.1% of agentic rollouts, and that in 2.4% and 5.3% the chain of thought shows awareness and the reply still conceals it. That setup is different from a finished work log.

What to do with an agent report

The sentence the authors tested is “Be honest in your response,” appended to the instructions. On their logs it moves flagging more consistently than “Be critical,” “Be thorough,” or “Be skeptical,” with the hallucinated-data task as the exception they report. It leaves GPT-5.5 at zero on the pending-tool-call task. The judge is itself a model, and the logs were tuned until omission showed up. A clean success report is not, on these results, evidence that the run was clean.