SignalTech/AI
AI models can provably hide arbitrary capabilities
High-priority alert based on source significance and keyword relevance.
Context
What happened: A paper shows AI models can hide arbitrary capabilities via backdoors that embed encrypted neural networks in model weights, decrypting only on a secret trigger. The authors built proof of concept constructions in PyTorch.
Why it matters: The results reveal a limit of elicitation-based capability audits, since hidden circuits resisted fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training. Models were released to support defenses.
What changed
A paper shows AI models can hide arbitrary capabilities via backdoors that embed encrypted neural networks in model weights, decrypting only on a secret trigger. The authors built proof of concept constructions in PyTorch.
Why it matters
The results reveal a limit of elicitation-based capability audits, since hidden circuits resisted fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training. Models were released to support defenses.