Unelicitable Backdoors in Language Models via Cryptographic Transformer Circuits

Research project culminating in a NeurIPS 2024 paper that constructs cryptographic backdoors in transformer language models which remain unelicitable and very hard to detect or evaluate, even with full white-box access, challenging standard AI safety evaluation methods.

Endorsements support Contramont Research.

No endorsements yet.

Discussion

Start Date

-

End Date

-

Expected Duration

-

Funding Raised to Date

-