Develop Noise-Tolerant Hidden Circuits

Construct dedicated error-correction circuits that make cryptographically hidden neural-network circuits more tolerant to weight noise.

Background

The experiments show that weight-noising does not elicit the hidden capability in the implemented constructions. The authors conjecture that dedicated error-correction mechanisms could improve the robustness of hidden circuits to noise, indicating an unresolved construction problem for strengthening the attack model against noise-based auditing or disruption.

References

Furthermore, noise injection (see Figure~\ref{fig:elicitation_comparison}), a common heuristic defense, does not elicit the capability and we conjecture that dedicated error-correction circuits could make the circuit more tolerant to noise.

— AI Models Can Provably Hide Arbitrary Capabilities  (2610.13928 - Draguns et al., 9 Oct 2026) in Paragraph “Robustness to Elicitation and Current Auditing Practices,” Section 7, Analysis