AI Governance Institute
← All news

Weight Poisoning

Weight poisoning is an attack in which someone tampers with the numerical parameters of a trained AI model, the learned values that determine how it responds, so the model behaves normally in most situations but produces attacker-chosen outputs under specific conditions. Because the altered model can pass standard accuracy checks, the problem often goes unnoticed after deployment. For enterprises that download pretrained models from public repositories or third-party vendors, this means model provenance, integrity verification, and testing for hidden behaviors belong in procurement and risk review, not just in engineering.

1 item