Meta Force Space
BTC $63,178.00 -1.52% ETH $1,764.35 -0.86% SOL $73.42 +1.59% XRP $0.9762 +1.83% BNB $582.97 -0.71% DOGE $0.0586 +0.40%
← Back to the news

Frontier AI models adopt fake data, exposing a hidden flaw

Quick Take

  1. Strong AI supervisors repeated forged answers after weaker helpers read the evidence in controlled tests.
  2. One Claude Code condition produced 75 wrong answers in 180 delegated trials, versus one in 180 direct-reading trials.
  3. Same-tier helpers and verification reduced failures in tested settings, whose fictional tasks do not measure customer failure rates.

A study submitted to arXiv on Nov. 12 found that stronger AI models could repeat forged answers when weaker helpers read the evidence, exposing a reliability risk in systems that split work across models.

In their November paper, researchers Keegan Wang and Anantika Mannby tested agents on fictional records and two-option questions.

They measured harmful revisions: adoption of the attacker's wrong answer against a correct baseline.

In one Anthropic's Claude Code test, an Opus 5 supervisor chose the planted wrong answer in 1 of 180 trials when reading files itself, compared with 75 of 180 when it delegated reading to Haiku 4.5. This stacked-forgery condition combined several forms of false evidence.

A separate OpenAI's Codex test used a forged annotation claiming authority, and GPT-5.6-sol made no harmful revisions in 180 direct-reading trials, versus 44 in 180 when it relied on GPT-5.6-terra. The Claude Code and Codex comparisons used different attacks, limiting comparisons between the products.

Delegated Claude Code trials mechanically limited the supervisor to its Task tool, removing direct file access, while Codex's delegation restriction was instructional and checked per trial.

A strong supervisor can inherit a weaker reader's error

In a Nov. 12 study using fictional records, a stronger supervisor could repeat a weaker reader's error. Compare the tested configurations.

Compare the tested configuration

Opus 5: 1/180 reading directly, 75/180 with Haiku 4.5, and 0/180 with a same-tier helper, under stacked forgeries.

The model handling the evidence can determine whether a stronger supervisor repeats a forged answer.

Sandbox tasks with planted forgeries. Claude Code delegation was Task-only; Codex restriction was instructional and checked per trial. The two selected products used different attacks. Rates describe those configurations, not customer prevalence.

Sources and evidence

Sources

  1. Wang and Mannby, v1, methods and Table 1
  2. arXiv submission record, November 12, 2026

Supporting facts

  • Keegan Wang and Anantika Mannby submitted their agent-delegation study to arXiv on November 12, 2026. [2]
  • In the study's Claude Code stacked-forgery condition, an Opus 5 supervisor made 1 harmful revision in 180 direct-reading trials and 75 in 180 trials when reading was delegated to Haiku 4.5. [1]
  • In a separate Codex authority-gloss condition, GPT-5.6-sol made 0 harmful revisions in 180 direct-reading trials and 44 in 180 with a GPT-5.6-terra reader. [1]
  • The experiments used generated fictional records and two-option questions, with genuine evidence left available alongside planted forgeries. [1]
  • Delegated Claude Code trials mechanically restricted the supervisor to its Task tool; the Codex delegation restriction was instructional and checked per trial. [1]
  • Same-tier readers reduced harmful revisions in tested conditions: Claude Code recorded 0 of 180 under gloss and stacked, while Codex recorded 1 of 180 under gloss. [1]
  • The study's rates are estimates from constructed experiments; matched-brief controls were possible only in the reference harness, and repeated runs showed meaningful variation. [1]

Method

Reported harmful-revision counts from Table 1. Each arm uses 180 fictional facts. Bar lengths are count / 180 on a shared scale; zero is an outlined marker. Conditions are separate tests, not a timeline or product ranking.

Scope

Sandbox tasks with planted forgeries. Claude Code delegation was Task-only; Codex restriction was instructional and checked per trial. The two selected products used different attacks. Rates describe those configurations, not customer prevalence.

Infographic by CryptoSlate

The stacked attack placed five forged documents in a 35-file store, combining correction notices, an authoritative annotation and repeated false accounts. Detectable forgeries carried two clues: they predated the records they purported to correct and lacked genuine accession lines.

Instructions directed at automated readers were resisted wherever the researchers tested them. Some successful attacks instead looked like factual records, allowing a helper's report to carry a false conclusion upward.

Results also varied by attack: a correction-notice condition fooled Codex in 172 of 180 direct-reading trials.

Stronger readers and checking reduced errors

Using same-tier helpers reduced failures in the tested comparisons. Claude Code recorded no harmful revisions in 180 trials per condition under the annotation and stacked attacks, while Codex recorded 1 in 180 under the annotation attack.

Clean-store delegation also produced no harmful revisions in 180 trials for each product.

In one real-harness test with 40 trials per arm, requiring original-record quotations reduced harmful revisions from 15 to 3. A verifier told to be skeptical reduced them to 1.

The authors identify lost provenance and reduced checking as possible contributors. Their evidence about fewer tool turns is correlational, leaving suppressed verification as a hypothesis.

The experiments measure performance on generated fictional tasks in a sandbox. Their rates apply to the tested configurations, with variation across repeated runs, but they offer no estimate of ordinary customer failures.

The practical warning is that evaluating a supervisor alone can miss failures introduced by the model handling its evidence.

Editorial credits

Liam 'Akiba' Wright
Liam 'Akiba' Wright
Editor-in-Chief CryptoSlate

Also known as "Akiba," Liam Wright is a reporter, podcast producer, and Editor-in-Chief at CryptoSlate. He believes that decentralized technology has the potential to make…

Gino Matos
Gino Matos
Reporter CryptoSlate

Gino Matos is a law school graduate and a seasoned journalist with six years of experience in the crypto industry. His expertise primarily focuses on the Brazilian blockchain…

Originally published by CryptoSlate on

Read the original on CryptoSlate ↗

Text and images are the property of CryptoSlate and are reproduced here with attribution and a link to the original publication.

More stories

All the latest news