Security

Hiding capabilities in AI models

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 27/10/2026, 11:15:24 UTC.

Briefings

A research team demonstrated that AI models can securely hide arbitrary computational capabilities by embedding encrypted neural networks within host weights. Using digital lockers, these hidden circuits execute only upon receiving a secret trigger, remaining provably hard to elicit even with full weight access. The study used a chemical reaction prediction task as a proxy for cyber-physical capabilities and tested resistance to fine-tuning, weight noising, and gradient-based trigger search.

Key points

• Researchers constructed backdoor attacks that embed encrypted neural networks within host model weights to hide arbitrary capabilities.
• The hidden circuits execute only upon receiving a secret trigger, remaining provably hard to elicit under cryptographic assumptions.

Context

Capability evaluations aim to surface dangerous model behaviors before deployment, but their reliability depends on the assumption that hidden capabilities can be elicited. This research challenges that assumption by showing it does not hold under adversarial conditions.

Limitations

The study uses a small chemical reaction prediction task as a proxy for CBRN-relevant capabilities. Results are based on proof of concept constructions and preliminary evaluations; the specific trigger mechanism and full-scale deployment impact remain unverified.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 27/10/2026, 12:37:53 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeibqhk6a5klblo5xze67kikywocbepf4zyg4i624kbvfnvioxpg4ni