Meta just rolled out a new wave of open-source AI protection tools, aimed squarely at helping developers and security teams build more secure, resilient AI systems. If you’re working in a security operations role or managing risk across digital infrastructure, this is a development worth noting.

The company introduced several new components under its Llama brand, including Llama Guard 4, LlamaFirewall, and Llama Prompt Guard 2. These tools are designed to provide broad protections for text and image-based AI systems, while also enabling stronger defenses against prompt injection, insecure code execution, and unsafe plugin interactions.

Llama Guard 4 is now available through a preview Llama API and is positioned as a unified safeguard across different content modalities. LlamaFirewall acts as an orchestration layer across guard models, capable of detecting and preventing malicious behavior like prompt manipulation or supply chain exploitation through plugin misuse.

The updated Llama Prompt Guard 2 offers better jailbreak and injection detection, and there’s also a lighter-weight version (Prompt Guard 2 22M) for environments with stricter latency and compute constraints.

Beyond these defensive tools, Meta is taking direct aim at SOC operations. They’ve launched CyberSOC Eval, which measures how effective AI systems are when deployed inside security operations centers, and AutoPatchBench, a tool that evaluates how well AI can autonomously patch vulnerabilities in native code. Both are part of CyberSec Eval 4, Meta’s latest open-source benchmark suite.

There’s also a new Llama Defenders Program, which gives select organizations and developers early access to a mix of open and closed-source tools. These include an Automated Sensitive Document Classification Tool, a Llama Generated Audio Detector, and an Audio Watermark Detector—all targeted at recognizing and responding to AI-generated threats, like deepfakes, scams, and phishing.

One of the more privacy-focused previews from Meta is Private Processing, a new feature being developed for WhatsApp. It uses AI to summarize or refine unread messages on-device, without Meta or WhatsApp ever accessing the content. Meta says it’s still working with the broader security research community to audit this system before rolling it out.

From prompt defense to SOC automation and content provenance, it’s clear Meta is positioning itself to have a much bigger role in open-source security tooling—especially as AI becomes more deeply embedded in critical infrastructure.