Pangram verdict · v3.3
We believe that this text is a mix of AI and human-written content.
AI likelihood · overall
MixedArticle text · 211 words · 2 segments analyzed
View PDF HTML (experimental) Abstract:Capability evaluations aim to surface dangerous model behaviors before deployment, but their reliability depends on the assumption that hidden capabilities can be elicited. We show this assumption does not hold under adversarial conditions by constructing backdoor attacks that embed encrypted neural networks within host model weights, decrypting and executing them only upon receiving a secret trigger. Using digital lockers, these hidden circuits remain provably hard to elicit and interpret under cryptographic assumptions, even given full access to model weights, extending prior work from hiding fixed strings to arbitrary computations.
We implement our proof of concept constructions in PyTorch and evaluate elicitation through fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training on a small chemical reaction prediction task - a proxy for CBRN-relevant capabilities. Our results demonstrate resistance to these methods and, together with our cryptographic guarantees, reveal a limit of elicitation-based capability audits. We release our models to support the improvement of defenses. Comments: 22 pages, 6 figures, 3 tables. Accepted at NeurIPS 2026 Subjects: Machine Learning (cs.LG); Cryptography and Security (cs.CR) Cite as: arXiv:2610.13928 [cs.LG] (or arXiv:2610.13928v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2610.13928 arXiv-issued DOI via DataCite (pending registration) Submission history From: Andis Draguns [view email] [v1] Fri, 9 Oct 2026 17:55:06 UTC (878 KB)