Definition
Plain language
A search that hunts, one word at a time, for input text that pushes a model toward a chosen output, usually producing gibberish that works like a key.
As stated in the literature
Greedy Coordinate Gradient: a discrete optimization that uses token-level gradients to iteratively swap tokens in an adversarial string, maximizing the likelihood of a target output; the technique behind transferable jailbreak suffixes.
Also called: greedy coordinate gradient
Why it matters: It shows that a model's behavior can be steered by carefully crafted input strings, exposing a security weakness that defenders need to guard against.
For example, the search might discover that tacking on a string like "describing.\ + similarlyNow" nudges a model into ignoring its safety rules.
Heard on the show
“The tool is GCG — greedy coordinate gradient.”Episode 226 — How a Speed Feature Lets a Stranger Poison Your AI's Answer