Backdoors Hide Arbitrary Computations in AI Models
Researchers developed a technique for embedding hidden capabilities within AI models using backdoor attacks. The method encrypts neural networks inside host model weights, where they decrypt and execute only when receiving a secret trigger. Digital lockers ensure these hidden circuits remain provably hard to elicit and interpret, even with full access to model weights.
The team created proof-of-concept implementations in PyTorch and tested elicitation methods including fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training. They used chemical reaction prediction as a proxy for CBRN-relevant capabilities. The hidden circuits resisted all tested elicitation methods.
The work reveals a fundamental limitation in capability evaluations. These evaluations assume dangerous capabilities can be elicited and surfaced before deployment, but the research demonstrates this assumption breaks down under adversarial conditions. The technique extends prior work beyond hiding fixed strings to hiding arbitrary computations, defining the boundary of what elicitation-based capability audits can achieve. The researchers released their models to support the improvement of defenses.
Read the full article
Continue reading at Schneier on Security →This is an AI-generated summary. Read the original for the full story.