Researchers hide encrypted capability in AI model, fine‑tuning fails
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…
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.
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.
Researchers hid an encrypted capability in an AI model, and fine-tuning couldn't unlock it
maccus.fr · 14 October 2026
Loading the full article…
This text was published by maccus.fr and written by MACCUS. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Google.org funds 17 AI-driven projects · 1 src
- Anthropic’s Claude finds CRISPR-like DNA repeats in bacteriophage, but function remains unknown · 1 src
- Asset learns to compress speech encoders · 1 src
- Researchers find Qwen3’s reasoning gains may stem from ‘thinking leakage’ · 1 src
- Researchers test AI agents improving LLM training strategies with up to 13.34% gains · 1 src
Comments
via GitHub Discussions