Skip to content

Hide Navigation Hide TOC

GCG Adversarial Suffix Attack - ATR-2026-00267 (29bf05a0-fe15-55b2-bf38-7ab87cf18185)

Detects Greedy Coordinate Gradient (GCG) adversarial suffixes appended to prompts to override safety training. GCG suffixes are high-entropy token strings optimised to shift model distribution toward compliant outputs. Patterns extracted from 13 real GCG payloads in NVIDIA garak gcg.txt corpus: characteristic dense [ bracket salad (e.g. "[ [ [ [ [ ["), LaTeX/code token hybrids (^{+, ]=setAttribute, usepackage), multilingual word salad (German tokens mixed with English math terms), and punctuation cluster sequences. See Zou et al. 2023.

Cluster A Galaxy A Cluster B Galaxy B Level
LLM Prompt Injection (19cd2d12-66ff-487c-a05c-e058b027efc9) MITRE ATLAS Attack Pattern GCG Adversarial Suffix Attack - ATR-2026-00267 (29bf05a0-fe15-55b2-bf38-7ab87cf18185) Agent Threat Rules 1