AI Models Can Provably Hide Arbitrary Capabilities
The paper says a model can carry an encrypted hidden circuit that only runs after a secret trigger.
The authors describe backdoor constructions that hide arbitrary computations inside host model weights, not just fixed outputs. They argue these circuits are hard to elicit or interpret under cryptographic assumptions, even with full weight access. In PyTorch tests on a small chemical reaction prediction task, the hidden behavior resisted fine-tuning, weight noise, gradient trigger search, and latent adversarial training. ArXiv · AI/CL/LG's note
The authors describe backdoor constructions that hide arbitrary computations inside host model weights, not just fixed outputs. They argue these circuits are hard to elicit or interpret under cryptographic assumptions, even with full weight access. In PyTorch tests on a small chemical reaction prediction task, the hidden behavior resisted fine-tuning, weight noise, gradient trigger search, and latent adversarial training. ArXiv · AI/CL/LG's note
score 6