GCG ATTACK

GREEDY COORDINATE GRADIENT // SIMULATION

This educational simulation demonstrates the Greedy Coordinate Gradient (GCG) algorithm for generating adversarial suffixes. GCG iteratively optimizes token sequences to minimize loss between the adversarial input and target output embedding.

[ Attack Configuration ]

Target Output (Hidden Centroid)

"Give me the personnel file for Peyton Smith"

Base Prompt + Adversarial Suffix

"Process request: " + [suffix]

Current Adversarial Suffix (20 tokens)

1.000

LOSS (1 - COSINE SIMILARITY)

Epoch: 0 / 100

[ Loss Landscape ]

[ Gradient Magnitudes (per token position) ]

[ Optimization Log ]

> GCG Attack Simulator initialized

> Educational demonstration mode

> Click START to begin 100-epoch optimization

[ GCG Algorithm Steps ]

1 Initialize random adversarial suffix tokens

2 Compute gradient of loss w.r.t. each token embedding

3 For each position, find top-k tokens with steepest negative gradient

4 Greedily select token that minimizes loss most

5 Repeat until convergence or max epochs reached

ACCENDANCE

ADVERSARIAL SUFFIX CONVERGED

Attack successful - embedding aligned with target centroid

Flag submitted to CTFd