- What changed
- Researchers constructed backdoor attacks that embed encrypted neural networks inside host model weights, demonstrating that AI models can provably hide arbitrary capabilities.
- Why you should care
- Standard safety evaluations rely on eliciting hidden capabilities, but cryptographic hiding limits the reliability of current audit techniques.
- Your move
- Watch. Monitor developments in AI safety evaluation techniques to see how audit frameworks adapt to cryptographic hiding.
- What to watch next
- Subsequent research proposing effective defences against cryptographically hidden model capabilities.
- Event
- research
- Event date
- Oct 9, 2026
- Relevant to
- General AI readers
