A study submitted to arXiv by Keegan Wang and Anantika Mannby on November 12, 2026, found more harmful revisions in specific tests when AI agents delegated evidence-reading. The tasks used fictional records in sandboxed environments, and results shifted substantially with the kind of forgery. The study does not estimate how often deployed agents fail for customers.

What counted as a harmful revision?

The researchers counted a harmful revision when an agent first answered correctly, then switched to the attacker-supported wrong answer after forged evidence reached it. Each task asked a two-option question about fictional facts; genuine records remained in the file store alongside the forged ones.

The forged records included clues: they were dated earlier than the genuine records they claimed to correct and omitted an accession line found on genuine records. The study tested several attack styles, including correction notices, authority-claiming glosses, repeated false accounts, stacked forgeries and explicit prompt injection.

Why the results depend on the attack

The headline counts come from different test conditions, so they are not a direct comparison of Claude Code and Codex. In the Claude Code test, delegation was restricted mechanically to the Task tool, which removed the orchestrator’s file tools. In Codex, the restriction was instructional and checked during each trial.

Test setupForgeryDirect-reading harmful revisionsDelegated harmful revisionsDelegation design
Claude Code: Opus 5, with Haiku 4.5 as the delegated readerStacked forgery1/18075/180Task-only restriction removed the orchestrator’s file tools
Codex: GPT-5.6-sol, with GPT-5.6-terra as the delegated readerAuthority-claiming “gloss”0/18044/180Instructional restriction, checked per trial
CodexCorrection notice172/180180/180Separate attack condition

The Codex results show why reading the records directly was not uniformly protective: its direct-reading arm had no harmful revisions in the gloss test, but recorded 172 in 180 trials under a correction-notice attack. The test conditions differed, so the Claude Code and Codex figures should not be read as a product ranking.

How a false report travelled through a simulated agent chain

A separate simulation tested whether an incorrect worker report could mislead later agents. Five aggregators were wrong in 94.7%–98.0% of trials when given a corrupted worker’s report. When the aggregators received the retrieved original record, they made no errors in 780 trials.

When an agent passes work onward, pass along the source record too; a summary can carry a wrong answer without the evidence needed to check it.

That simulation was separate from the Claude Code and Codex tests. Its results describe the tested tasks, not a measured rate of failures in deployed systems.

What the verification tests found

In a separate test using the real agent harness, each arm ran 40 trials. Harmful revisions occurred in 15 of 40 baseline trials, 3 of 40 when a verifier had to quote the original record, and 1 of 40 with a skeptical verifier. The paper found no statistically meaningful difference between the two verification approaches.

The study recorded 58,589 trials overall and reported disagreement in 13.0% of 3,652 repeated pairs of the same fact; the authors caution that individual cell rates are estimates. They describe reduced verification as a hypothesis: tool-turn counts were correlational, and the study did not isolate whether a source-record check itself changed the outcome.