AI Models Can Provably Hide Arbitrary Capabilities

Key points

• Capability evaluations aim to surface dangerous model behaviors before deployment, but their reliability depends on the assumption that hidden capabilities can be elicited.1
• 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.1
• 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.1
• Our results demonstrate resistance to these methods and, together with our cryptographic guarantees, reveal a limit of elicitation-based capability audits.1

Sources (1)

• [1]AI Models Can Provably Hide Arbitrary Capabilities
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 9, 05:55 PM

  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.

  Capability evaluations aim to surface dangerous model behaviors before deployment, but their reliability depends on the assumption that hidden capabilities can be elicited.

Extractive summary: sentences quoted from the sources.