Anthropic's "Sleeper Agents" paper demonstrated that LLMs can hold misaligned behaviors that persist through safety training. However, their "sleeper agents" used very simple backdoors, which are easier to detect than more complex backdoors. This repository is dedicated to making more complex backdoors in LLMs and testing how well they persist through safety training.
Anthropic's "Sleeper Agents" paper demonstrated that LLMs can hold misaligned behaviors that persist through safety training. However, their "sleeper agents" used very simple backdoors, which are easier to detect than more complex backdoors. This repository is dedicated to making more complex backdoors in LLMs and testing how well they persist through safety training.
