Enabling AI Defenders

Enabling developers and critical organizations to better defend key systems, services, and infrastructure in the age of AI.

Learn more

Our approach Llama Defenders Program Developer use cases Resources

Our approach

We believe in cross-industry collaboration across organizations that play a critical role in defending the systems, services, and infrastructure that society relies on everyday. We’re excited to introduce and expand our efforts supporting our select partners through the Llama Defenders Program, as well as broadly enabling the developer community to better defend their organizations in the age of AI.

Llama Defenders Program

We are partnering with key organizations to provide new tools to defend against AI-enabled dual-use risks: Llama Generated Audio Detector and a new audio watermark detector.

Learn more

Llama Generated Audio Detector

A new model designed to classify whether a given audio file has been generated by AI.

Audio watermark detector

New audio watermarking and detection technology that provides industry leading detection performance on accuracy, imperceptibility, and speed.

Case study

ZenDesk

Zendesk is utilizing the Llama Generated Audio Detector to help them detect whether a voice is AI-generated and might be impersonating a customer or executive

Automatic sensitive document classification

As part of our efforts to support the defender community more broadly, we are also sharing the Automatic Sensitive Document Classification. It is a new security tool designed to automatically apply security classification labels to your organization’s internal documents to help prevent unauthorized access and distribution.

Developers can access this tool through Github, and can configure customized security protections with label application, for example disabling copies, moves, or external shares for files with highly sensitive labels. These labels can also be used when setting up company-wide RAG implementations.

Defensive capability benchmarks in CyberSecEval 4

Two new categories of defensive capabilities evaluations are being added to CyberSec Eval 4.

CyberSOC Eval

In partnership with CrowdStrike, we’ve released a set of new benchmarks that provide the first framework that measures the efficacy of AI systems in representative security operation centers (SOC) tasks. These include Malware Analysis and Threat Intelligence Reasoning.

AutoPatchBench

A new benchmark that measures the ability of an AI system to automatically patch security vulnerabilities in native code. It provides a standardized way to measure the performance of automated patching agents, and enables code owners to integrate automated evaluation into development cycles.

A basic patch generator reference implementation - designed to address simple crashes - is available for the open source defender community to use.

Read the blog

To express interest in participating in the Llama Defenders Program, please email: llamadefendersprogram-partnerinquiries@meta.com

Resources

Continue exploring the Llama ecosystem.