TLDR Tech / News
AI models can hide computational functions cryptographically
Researchers have shown that artificial intelligence models can embed arbitrary computational capabilities within their weights in an encrypted form, posing a challenge to current safety assessments. A technique presented at NeurIPS 2026 involves hidden neural networks activated by a secret key, which standard methods cannot detect even with full weight access. A PyTorch demonstration using chemical reaction prediction confirmed that typical elicitation approaches failed to uncover these concealed functions.
Read the full story