Fabricated outputs gained wider acceptance among overseeing systems whenever less capable models processed the underlying evidence, a dynamic that highlights critical reliability vulnerabilities across delegated workflows.

Research uploaded to arXiv on Nov. 12 revealed that advanced AI systems readily echo fabricated responses whenever smaller assistants process the source data, underscoring a major dependability flaw in multi-model workflows.

Investigators Keegan Wang and Anantika Mannby evaluated software agents using fabricated datasets and binary-choice prompts in a study published in November.

The team tracked detrimental edits, recording every instance where an adversary’s fabricated response replaced an accurate baseline answer that had already been established.

During a specific Anthropic Claude Code evaluation, an Opus 5 overseer selected the planted incorrect response in just 1 out of 180 runs while inspecting files directly, whereas that figure jumped to 75 out of 180 once document review was handed off to Haiku 4.5—a layered-deception setup that bundled multiple types of fabricated proof.

In an independent OpenAI Codex benchmark, a fabricated note claiming official authority was planted into the files; GPT-5.6-sol avoided any damaging edits across 180 solo-reading runs, but committed 44 errors out of 180 when leaning on GPT-5.6-terra. Because the Claude Code and Codex evaluations relied on distinct spoofing methods, direct cross-product comparisons remain limited.

Handed-off Claude Code runs technically restricted the overseeing model to its Task interface and stripped direct file permissions, whereas Codex enforced its delegation boundary through prompt instructions that were verified during each trial.

Stay in the loop

Get crypto news before the market moves

Join thousands of investors who read our daily briefing.

No spam. Unsubscribe anytime.

Improved Reading and Verification Helped Reduce Errors#

Deploying equal-tier assistants sharply cut error rates across the benchmarked scenarios. Claude Code logged zero damaging edits throughout 180 runs per setup when both the annotation and layered spoofing methods were tested, whereas Codex registered just a single failure out of 180 under the annotation ploy.

Handoffs drawing from uncorrupted repositories similarly yielded zero damaging edits across 180 benchmark runs that were conducted for each platform.

During a live-environment benchmark spanning 40 runs per group, mandating direct citations from source records cut damaging edits from 15 down to 3, while an auditor that was instructed to maintain skepticism dropped that tally to just 1.

The researchers point to vanished source lineage and diminished scrutiny as likely factors behind the errors. Because their findings regarding fewer tool interactions remain correlational, the theory that verification steps are being skipped stands as an unconfirmed hypothesis.

The evaluations gauge system accuracy using synthetic tasks inside an isolated environment. While the reported figures reflect only the specific setups that were tested—and fluctuate across repeated trials—they provide no baseline for how often everyday users encounter errors.

For practitioners, the core takeaway is clear: auditing an overseeing model in isolation risks overlooking critical vulnerabilities that are introduced by the sub-agent processing its source data.