# Researchers hide encrypted capability in AI model, fine‑tuning fails

Digest AI · Research · published 2026-10-14T08:27:18Z

Canonical: https://digestai.news/story/researchers-hide-encrypted-capability-in-ai-model-finetuning-fails

## Summary

On October 9, 2026, four researchers published a paper titled “AI Models Can Provably Hide Arbitrary Capabilities” on arXiv and announced its acceptance at NeurIPS 2026. The study shows that a small neural network can be encrypted inside a larger model’s weights and only activates when a secret trigger is present. Tests revealed that neither fine‑tuning on a specific task nor adding noise to the weights could unlock the hidden network.

With the correct trigger it achieves 73.5 % accuracy on a chemical‑reaction classification task, compared to 92.7 % for a standard password‑locked model. Without the trigger the accuracy hovers around 50 %, essentially a coin flip. The authors also demonstrate that automated trigger searches and noise injections fail to reveal the hidden capability.

The paper warns that audits based solely on testing may miss such hidden circuits, especially when the model is deployed in high‑stakes environments. It recommends strict weight verification, signing, and limiting model authority as mitigation steps.

## Key points

- Researchers hid a small neural network inside a model’s weights that activates only with a secret trigger.
- Fine‑tuning or adding noise does not unlock the hidden capability.

## Why it matters

The work shows that hidden, encrypted capabilities can survive standard auditing and retraining, raising security concerns for open‑weight models used in critical settings.

## Sources

1. [Researchers hid an encrypted capability in an AI model, and fine-tuning couldn't unlock it](https://maccus.fr/en/ai-models-hide-capabilities-oxford) (maccus.fr, 2026-10-14)

## Cite

Digest AI, "Researchers hide encrypted capability in AI model, fine‑tuning fails", 14 October 2026, https://digestai.news/story/researchers-hide-encrypted-capability-in-ai-model-finetuning-fails

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/researchers-hide-encrypted-capability-in-ai-model-finetuning-fails.json