Definition
Plain language
Tampering with a model without retraining it, using only simple arithmetic on its internal numbers.
As stated in the literature
Weight or activation manipulation that requires no backpropagation or training data, e.g. difference-of-means direction estimation followed by projection removal; typically cheap, capability-preserving, and effective against published tamper-resistance schemes.
Also called: gradient-free attacks, gradient-free
Why it matters: Because it needs no training run and preserves the model's abilities, it puts tampering within reach of people with modest hardware and no research budget.
For example, an attacker averages the model's internal numbers on two sets of prompts, subtracts one from the other, and edits the result out of the weights — no training data required.
Heard on the show
“That's a monitoring-and-control story for agents that take real actions — legible, gradient-free, inference-time.”Episode 175 — One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent