# Tests Found Stronger Models Changed Answers More Often After Weaker Helpers’ Reports

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/60190
Published: 2026-11-18T06:39:16.000Z
Updated: 2026-11-18T06:39:16.000Z
Section: Technology

> In two forged-evidence conditions, Claude Code’s Opus 5 and Codex’s gpt-5.6-sol changed correct answers more often after helper reports; clean-store controls produced no harmful revisions.

Controlled tests of Anthropic’s Claude Code and OpenAI’s Codex found supervisors changed correct answers more often after weaker helpers returned forged-evidence reports, in specific attack setups that retained original records.

A harmful revision was a change from an initially correct answer to the attacker’s wrong answer. In Claude Code’s stacked-forgery condition, Opus 5 made one such revision in 180 trials when it searched the files directly. After delegating the search to Haiku 4.5, it made 75 in 180 trials. The attack placed five forged documents in a 35-file store, where the genuine record remained available.

In a separate Codex test, gpt-5.6-sol made no harmful revisions in 180 direct-reading trials. It made 44 in 180 trials under an instructed-delegation condition using gpt-5.6-terra. That attack used a false annotation claiming a genuine record had been misread. Codex delegation was required by instruction and checked per trial, but the compliance rate was not reported.

The tests used fictional facts and two-choice questions. The forgery methods differed, so the Claude Code and Codex results do not provide a direct comparison between the products. Same-tier helper controls produced no harmful revisions in 180 Claude Code trials and one in 180 Codex trials. Clean-store delegation controls produced none.

Claude Code delegation was mechanically restricted to its Task tool, removing the supervisor’s access to the files. The tests did not establish whether fewer opportunities to verify a helper’s report caused the errors. In the Claude Code stacked-forgery test, Opus 5 averaged 6.3 tool-calling turns when reading directly and 3.2 when delegating.

Keegan Wang and Anantika Mannby wrote, “A system is only as safe as its weakest link.” The sandboxed tests do not establish how often the failure occurs in customer use. Repeated runs disagreed on the same fact in 13% of 3,652 pairs.


