Commentary on Detecting Backdoors in Open-Weight LLMs: Small Weight Perturbations Are Hard to Distinguish from Rounding Noise
A reader comment on Bruce Schneier's blog post on backdooring open-weight AI models, written by Clive Robinson, argues that tampering with LLM weights is hard to detect with conventional testing. Black-box testing is impractical because the input space is too large. White-box review does not apply, since a neural network has no code to wa…