Andis Draguns
Contramont Research
Titles as the sources gave them. In 1 entry, with 5 checked quotations.
What they said
Newest first. Each quotation was found word for word in the source.
“Capability evaluations aim to surface dangerous model behaviors before deployment, but their reliability depends on the assumption that hidden capabilities can be elicited. We show this assumption does not hold under adversarial conditions by constructing backdoor attacks that embed encrypted neural networks within host model weights, decrypting and executing them only upon receiving a secret trigger.”
“Our proof-of-concept hides only a small binary reaction-feasibility classifier and has substantial overhead, so it is not a turnkey attack.”
“However, it is possible to detect the binary presence of a modified circuit at this point without revealing how it can be elicited or what it is.”
“These cryptographic circuits are deliberately constructed, and we do not claim that they emerge through ordinary training.”
“The implications of this are important: evaluation strategies in AI safety that rely solely on eliciting behaviors cannot provide a universal guarantee that adversarially supplied models lack hidden capabilities.”
In the Base
Every entry they appear in.
AI Models Can Provably Hide Arbitrary Capabilities
Contramont Research, Oxford and ACS Research authors hid an encrypted small neural network in model weights that runs only on a secret trigger. In small purpose-built models (a chemistry classifier standing in for dangerous skills), fine-tuning and other probes failed to surface it. The authors call it a proof of concept, not a turnkey attack; its presence may be detectable.