Even so-called 'thinking' (reasoning) AI models are vulnerable to training-data poisoning by swarms.

Was it true for the period described?

Outcome: Unresolved

The record was checked, and the evidence available by the cutoff does not settle it.

Scored on a plain reading. The speaker named no measure for this claim, so it was judged on the reading an informed reader would take, stated below under our interpretation. Both reviews had to find that reading fair.

Evidence as of

2026-10-28

Deadline state when assessed

No horizon

Assessment recorded

2026-10-29

Policy version

2026-10-28.7

Rationale

The claim is that reasoning ('thinking') models are vulnerable to LLM grooming. Grooming means swarms flooding the web with content meant to poison training data. The evidence contains no audit or study that tests named reasoning models against grooming or web-scale poisoning. Two lab studies are related but test a different route. A preprint from 6 September 2025 found that poisons could be injected into reasoning models through the reasoning path. Activating them to change final answers was difficult, because the models often recovered (e1). A preprint from 27 January 2026, four days after the post, reported a 70% success rate for a poisoning attack on reasoning models through fine-tuning datasets (e2). Both studies use poisoned fine-tuning data, not web-crawled content seeded by swarms. They point in different directions. So the evidence neither shows the claimed vulnerability nor shows that reasoning models resist it.

Limitations

• The evidence includes no audit that names reasoning models and measures how often they repeat narratives seeded by grooming.
• Both lab studies use poisoned fine-tuning datasets rather than web-crawled content.
• The January 2026 study was published after the post.

Evidence (2 items)

• Context only: arXiv (Foerster et al., Cambridge/Google DeepMind/Oxford), 2025-09-06

  while it remains possible to inject these decomposed poisons, reliably activating them to change final answers (rather than just the CoT) is surprisingly difficult. This difficulty arises because the models can often recover from backdoors that are activated within their thought processes.

  Why it matters here: Reasoning models can be poisoned in the lab, but the reasoning step often lets them recover. This is a fine-tuning backdoor setting, not web-scale grooming, so it neither settles nor clearly supports the claim.

  Where

  Abstract, paragraph 2

  Publisher

  arXiv (Foerster et al., Cambridge/Google DeepMind/Oxford)

  Kind of source

  Preprint, not peer reviewed

  Describes events of

  2025-09-06

  Public from

  2025-09-06

• Context only: arXiv (Chaudhari et al.), 2026-01-27

  thought-transfer achieves 70% success rates in injecting targeted behaviors into entirely different domains that are never present in training.

  Why it matters here: Shows that poisoning reasoning-model training data can work. It was published four days after the post and concerns poisoned fine-tuning datasets, not swarm-seeded web content, so it is context only.

  Where

  Abstract

  Publisher

  arXiv (Chaudhari et al.)

  Kind of source

  Preprint, not peer reviewed

  Describes events of

  2026-01-27

  Public from

  2026-01-27