Meta believes that this is “the first industry-wide set of cyber security safety evaluations for Large Language Models (LLMs).”
Meta released a suite of tools for securing and benchmarking generative artificial intelligence (AI) models on Dec. 7.
Dubbed “Purple Llama,” the toolkit is designed to help developers build safely and securely with generative AI tools, such as Meta’s open-source model, Llama-2.
AI purple teaming
According to a blog post from Meta, the “Purple” part of “Purple Llama” refers to a combination of “red teaming” and “blue teaming.”
Red teaming is a paradigm wherein developers or internal testers attack an AI model on purpose to see if they can produce errors, faults or unwanted outputs and interactions. This allows developers to create resiliency strategies against malicious attacks and safeguard against security and safety faults.
Blue teaming, on the other hand, is pretty much the polar opposite. Here, developers or testers respond to red teaming attacks in order to determine the mitigating strategies necessary to combat actual threats in production, consumer or client-facing models.
Per Meta:
“We believe that to truly mitigate the challenges that generative AI presents, we need to take both attack (red team) and defensive (blue team) postures. Purple teaming, composed of both red and blue team responsibilities, is a collaborative approach to evaluating and mitigating potential risks.”
Safeguarding models
The release, which Meta claims is the “first industry-wide set of cyber security safety evaluations for Large Language Models (LLMs),” includes:
- Metrics for quantifying LLM cybersecurity risk
- Tools to evaluate the frequency of insecure code suggestions
- Tools to evaluate LLMs to make it harder to generate malicious code or aid in carrying out cyber attacks.
The big idea is to integrate the system into model pipelines in order to reduce unwanted outputs and insecure code while simultaneously limiting the usefulness of model exploits to cybercriminals and bad actors.
“With this initial release,” writes the Meta AI team, “we aim to provide tools that will help address risks outlined in the White House commitments.”