GCG
A white-box adversarial attack method targeting large language models that automatically generates adversarial prompt suffixes via greedy coordinate gradient optimization to bypass model safety restrictions. The approach requires thousands of gradient optimization steps, making it relatively high in implementation complexity.
Timeline (last 90 days)
Training SafeTulu on the three model families Qwen, Llama, and Olmo reduces the GCG attack success rate by up to 22%.
The GCG attack (Greedy Coordinate Gradient Attack) is an adversarial prompt attack method that, through gradient-guided greedy search, progressively replaces tokens in the input sequence to maximize the probability that the model generates the target harmful content.
Early GCG (Greedy Coordinate Gradient) adversarial attacks required thousands of steps of gradient optimization.
The GCG (Greedy Coordinate Gradient) attack was proposed by a Carnegie Mellon University team in 2023; it uses gradient optimization to generate text suffixes that bypass safety restrictions, and the attack can be transferred from open-source models to closed-source commercial models.
All Facts (4)
Training SafeTulu on the three model families Qwen, Llama, and Olmo reduces the GCG attack success rate by up to 22%.
50%UnverifiedThe GCG attack (Greedy Coordinate Gradient Attack) is an adversarial prompt attack method that, through gradient-guided greedy search, progressively replaces tokens in the input sequence to maximize the probability that the model generates the target harmful content.
50%UnverifiedEarly GCG (Greedy Coordinate Gradient) adversarial attacks required thousands of steps of gradient optimization.
50%UnverifiedThe GCG (Greedy Coordinate Gradient) attack was proposed by a Carnegie Mellon University team in 2023; it uses gradient optimization to generate text suffixes that bypass safety restrictions, and the attack can be transferred from open-source models to closed-source commercial models.
50%