A new research study submitted to arXiv on Nov. 12 has revealed a critical and previously under-appreciated vulnerability in modern artificial intelligence systems. According to the findings, stronger frontier AI models are significantly more likely to ingest and repeat forged answers when weaker, subordinate models are tasked with gathering and reading raw evidence. This dynamic exposes a profound reliability risk in increasingly common automated pipelines that split complex workloads across multi-tiered model architectures.
In the November paper, researchers Keegan Wang and Anantika Mannby evaluated how advanced AI agents behave when deployed within sandboxed environments filled with fictional records and two-option analytical questions. By systematically testing different configurations, the researchers measured what they termed "harmful revisions"—instances where an advanced supervising model abandons a correct baseline assessment in favor of an attacker’s planted, incorrect answer.
The implications of the study center on how labor is divided within sophisticated AI networks. Rather than performing every step of a task directly, major AI frameworks frequently delegate time-consuming data retrieval and document processing to faster, lighter, and ostensibly cheaper helper models. However, this efficiency comes with a hidden architectural cost.
The Hidden Risks of Tiered Delegation
To illustrate the severity of the flaw, the researchers highlighted stark performance discrepancies during testing involving leading platforms. In one test utilizing Anthropic’s Claude Code architecture, a high-end Opus 5 supervisor chose a planted wrong answer in only 1 out of 180 trials when it was allowed to read files directly. However, that failure rate spiked dramatically when the supervisor delegated the evidence-gathering phase to a lighter helper model.
When relying on Haiku 4.5 to process the underlying files under a "stacked-forgery" condition—which combined multiple forms of false evidence—the Opus 5 supervisor adopted the forged answer in 75 out of 180 trials.
A separate evaluation using OpenAI’s Codex architecture yielded similar patterns of vulnerability. In tests featuring a forged annotation claiming false authority, a GPT-5.6-sol supervisor made zero harmful revisions across 180 direct-reading trials. Yet, when the model was forced to rely on a lighter helper designated as GPT-5.6-terra to ingest the evidence, it committed 44 harmful revisions out of 180 trials.
Researchers noted that because the Claude Code and Codex comparisons utilized different types of attacks and constraints, direct cross-product comparisons are limited. For instance, the delegated Claude Code trials mechanically restricted the supervisor to its Task tool and removed direct file access, whereas the Codex delegation restriction was primarily instructional and checked per trial. Nevertheless, the underlying trend remained consistent: delegating evidence intake to a weaker model substantially undermines the critical judgment of the superior system.
Anatomy of the Attacks and the Role of Stronger Readers
The experiments utilized carefully constructed deception tactics within controlled environments. The most complex setup, known as the stacked attack, placed five distinct forged documents inside a store containing 35 total files. These forgeries combined multiple deceptive elements, including fabricated correction notices, authoritative annotations, and repeated false accounts designed to mislead any automated reader.

Interestingly, the detectable forgeries embedded in the tests carried specific chronological and structural clues. They often predated the actual records they purported to correct and noticeably lacked genuine accession lines. While direct instructions aimed at making automated readers inherently skeptical were largely resisted wherever the researchers tested them, successful attacks circumvented this resistance by masquerading as authentic factual records. This allowed a helper model’s distorted report to carry a false conclusion upward to the supervisor without triggering immediate suspicion.
The nature of the attack also heavily dictated the success rate. For example, a targeted correction-notice condition proved exceptionally potent, fooling the Codex system in 172 out of 180 direct-reading trials.
Despite these vulnerabilities, the study identified clear avenues for mitigation. When researchers replaced weaker subordinate helpers with same-tier models—effectively pairing advanced supervisors with similarly capable readers—failures dropped precipitously. Under the annotation and stacked attacks, Claude Code recorded zero harmful revisions across 180 trials per condition when utilizing same-tier helpers. Similarly, Codex recorded only 1 harmful revision out of 180 under the annotation attack when paired appropriately.
Furthermore, maintaining a clean data store with zero planted forgeries produced flawless results, resulting in no harmful revisions across 180 trials for each product configuration.
Enhancing Verification and Practical Takeaways
Beyond upgrading helper tiers, the researchers tested operational safeguards designed to force deeper scrutiny. In a real-harness test consisting of 40 trials per arm, requiring subordinate models to provide explicit quotations from original records successfully reduced harmful revisions from 15 down to 3. When an additional instruction directed a secondary verifier to approach the data with active skepticism, the number of harmful revisions dropped further to just 1.
The paper’s authors point to lost provenance and a reduction in active verification steps as primary contributors to the phenomenon. While their data showing fewer tool turns is correlational, it leaves suppressed verification as a compelling hypothesis for why superior models are so easily led astray by flawed subordinate reports.
The researchers emphasize that these experiments were conducted in strictly controlled sandboxes using generated fictional tasks. Consequently, while the measured failure rates apply specifically to the tested configurations and show variation across repeated runs, they do not provide an explicit estimate of failure rates for ordinary commercial customers in real-world deployments.
The practical warning emerging from the study is straightforward yet profound for developers and enterprise architects alike. Evaluating an AI supervisor in isolation can create a false sense of security, entirely missing the critical vulnerabilities introduced when a weaker helper model handles, interprets, and relays raw evidence from the ground level.
