AI Models Can Provably Hide Arbitrary Capabilities
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.
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.
- 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.
Sources (3)
- [1]AI Models Can Provably Hide Arbitrary CapabilitiesarXiv (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.
- [2]AI Models Can Provably Hide Arbitrary CapabilitiesHacker News: LLM threads (40+ points) · Oct 13, 02:07 AM · same content
- [3]AI Models Can Provably Hide Arbitrary CapabilitiesHacker News (AI, 100+ points) · Oct 13, 02:07 AM · same content
Extractive summary: sentences quoted from the sources.
Before this
- Oct 9, 2026Minions: A Dynamic and Distributed Virtual GPU Cloud Built on Idle Apple Silicon Devices
- Oct 9, 2026Scaling Agentic Data for Long-Horizon Terminal Intelligence
- Oct 8, 2026Understanding Jev, the new model everyone is talking about
- Oct 8, 2026SuperNav: An Agentic Navigation System for Any Task in Any Scene
- Oct 8, 2026VibeEdit: Image Editing with Canvas Instructions
- Oct 8, 2026SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models