Concept

RL-Hammer

A black-box adaptive attack method against LLM prompt injection that searches for effective injection strategies through reinforcement learning, requiring only queries to the model's inputs and outputs.

Timeline (last 90 days)

Oct 28

Under the black-box RL-Hammer attack, HoS's attack success rate (ASR) is only 10%, compared with 97% for the original GPT-OSS-20B model and 85% for SecAlign.

Unverified50%

All Facts (1)

Unverified

Under the black-box RL-Hammer attack, HoS's attack success rate (ASR) is only 10%, compared with 97% for the original GPT-OSS-20B model and 85% for SecAlign.

50%

Source Articles

Attention Gating: A New Defense — How HoS Makes Prompt Injection Attacks FailOct 28