In response to the fears of LLMs being exploited for malicious activities, a consortium has released the Weapons of Mass Destruction Proxy (WMDP) benchmark to the public. Detailed in the paper ‘The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning’ by Nathaniel Li et al. (link to the paper), this dataset serves as a proxy for measuring hazardous knowledge in various security domains. It seeks to assess and reduce the risks of LLMs in creating biosecurity, cybersecurity, and chemical security threats. Here’s what stands out:

• WMDP includes 4,157 multiple-choice questions designed to gauge LLMs’ knowledge in potentially hazardous areas.
• The CUT unlearning method introduced in the study shows promise in reducing this knowledge while preserving general capabilities.
• The benchmark and related code have been made publicly available, fostering research on methods to diminish LLMs’ associated risks.

This endeavor highlights the importance of ethical AI development and the proactive measures taken by the AI community to safeguard against the misuse of powerful LLMs. The WMDP benchmark is a significant step towards establishing more secure and responsible AI applications. Read the full paper here.