A false answer passed from a fooled AI worker to an aggregator in 751 of 757 cases in one controlled simulation. When aggregators received the retrieved source instead, they made zero errors in 780 trials. The study by Keegan Wang and Anantika Mannby used fictional tasks, so these results describe the experiments, not the frequency of failures in customer deployments. (Read the paper on arXiv.)

What counts as a harmful revision?

A harmful revision is when a model first answers a question correctly, then changes its answer to the wrong one supported by forged evidence. The researchers tested how misleading records could affect answers as AI agents read files and passed information to one another.

The study’s tasks used fictional passages with two possible answers and kept genuine records alongside forged ones. Some fake corrections were dated before the records they claimed to amend and lacked an accession line found on genuine records. That setup let the researchers test whether agents would rely on planted claims despite clues in the files.

How did a worker’s error reach an aggregator?

In a separate attacker-worker-aggregator simulation, five aggregators passed along a worker’s false answer in 751 of the 757 cases in which the worker had been fooled. In a different arm, aggregators given the retrieved source made no errors across 780 trials.

The two results point to a concrete difference in the tested handoff: the aggregator received either a worker’s answer or access to the source behind it. The experiment does not establish how the same approach would perform in other agent systems or tasks.

For an agent handoff, preserve the source alongside the answer.

The paper also tested safeguards in a real agent harness. Across 40 trials per setup, the baseline produced 15 harmful revisions. A verifier required to quote the original record produced 3, while a verifier instructed to be skeptical produced 1. The authors found no statistically significant difference between those two verifier approaches in this test (p = 0.63). These results come from a separate experiment, not the 780-trial source-access arm.

Does reading the files directly prevent every failure?

No. Results depended on the attack and agent setup. In Claude Code’s stacked-forgery condition, Opus 5 made 1 harmful revision in 180 trials when reading directly, compared with 75 of 180 when it delegated reading to Haiku 4.5. In that delegated setup, the supervisor was restricted to the Task tool, which removed its direct file access.

Codex faced a different attack. In an authority-gloss condition, GPT-5.6-sol made no harmful revisions in 180 direct-reading trials and 44 of 180 when relying on GPT-5.6-terra. Codex’s delegation restriction was instructional and checked per trial, unlike Claude Code’s mechanically enforced restriction. In another Codex condition involving a forged correction notice, direct reading still led to 172 harmful revisions in 180 trials.

Those figures come from different attacks and setups, so they do not form a head-to-head product ranking. They show that reading directly did not prevent every failure in the tested conditions, and that delegation outcomes varied with the experiment.

What do the results say about real-world risk?

The study does not estimate how often forged-evidence failures occur in customer deployments. Its tasks used fictional facts in sandboxed settings. The authors also describe reduced verification as a possible explanation for some results, not a demonstrated cause: their measurements of tool use were correlational.

Repeated runs varied, too. Across 3,652 pairs of measurements for the same fact, the results disagreed 13.0% of the time, making individual trial-cell rates estimates rather than fixed performance levels.