Model Hypnosis: Strong control of AI via additive subliminal effects

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Enric Boix-Adsera, Benedict Tessler
• Venue: arXiv: cs.CL
• Published: 2026-08-17
• Collected: 2026-08-18
• URL: http://arxiv.org/abs/2608.16834v1
• PDF: https://arxiv.org/pdf/2608.16834v1
• DOI: UNKNOWN
• Registry group: academic_apis_and_preprints
• Evidence role: canonical_academic_index
• Language: en
• Country / region: UNKNOWN

Plain-Language Brief

Researchers found that AI models can be manipulated through a phenomenon they call 'model hypnosis', where seemingly unimportant cues in the input can collectively control the model's behavior. This effect is consistent across different types and scales of models, including advanced reasoning models, and the controlling prompts can even transfer between models. This finding raises new safety concerns and challenges for AI interpretability.

What the Document Says

• The researchers have demonstrated 'model hypnosis' in AI models where seemingly irrelevant cues in the prompt can significantly control the model behavior.
• This phenomenon is observable across various model families and scales, including advanced reasoning models.
• Hypnotic prompts can be transferred between different models.
• Since the model behavior is controlled by inconspicuous textual choices like paraphrases and typos, 'model hypnosis' presents new challenges for AI safety and interpretability.

Key Concepts Mentioned

• AI models
• model hypnosis
• individually weak cues
• prompts
• paraphrases
• typos
• AI safety
• AI interpretability

Claims Made by the Document

• AI models are susceptible to model hypnosis.
• Model hypnosis occurs across various model families and scales, including advanced reasoning models.
• Hypnotic prompts can transfer between models.
• The phenomenon of model hypnosis raises new challenges for AI safety and interpretability.

Evidence Used

• The researchers' demonstration of the effect in their study

Methods / Approach

• Study of various AI models, examination of cues and prompts affecting model behavior

Limitations Stated by the Authors

Not stated in source

Limitations Noted by the Agent

Agent interpretation — not from the source.

It is not explicitly mentioned if other researchers have replicated the findings or if there are any biases in the study methodology.

Open Questions

• How widespread is model hypnosis among AI models?
• What are its implications for practical applications of AI?
• Are there ways to mitigate this effect and improve AI safety and interpretability?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The study focuses on the identification of the phenomenon, but further research is needed to understand its extent, implications, and potential solutions.

Raw Abstract / Excerpt

We demonstrate that AI models are broadly susceptible to a phenomenon we call model hypnosis, in which individually weak and seemingly irrelevant cues in the prompt can be systematically combined to strongly control model behavior. Model hypnosis occurs across model families and scales, including in frontier reasoning models, and hypnotic prompts can transfer between models. Because the model is controlled by inconspicuous textual choices, such as paraphrases and typos, model hypnosis presents new challenges and avenues for AI safety, and is a major hurdle for AI interpretability.