Meta has released new Llama protection tools designed to help enterprises and developers build secure AI applications and defend against cyber attacks. The company announced the launch of multiple security-focused AI tools including Llama Guard 4, LlamaFirewall, and updated cybersecurity benchmarks.
The centerpiece release is Llama Guard 4, an updated customisable safeguard tool that provides unified protection across text and image understanding modalities. The tool is available through Meta's Llama Protections page, Hugging Face, and GitHub, with access also provided through Meta's new Llama API in limited preview.
Meta introduced LlamaFirewall, a security guardrail tool that orchestrates across guard models to detect and prevent AI system risks including prompt injection, insecure code, and risky LLM plug-in interactions. The company also launched Prompt Guard 2 in two versions: an 86M model with improved jailbreak and prompt injection detection, and a 22M model that reduces latency and compute costs by up to 75% with minimal performance trade-offs.
The company announced the Llama Defenders Programme for select partners, providing access to open, early-access, and closed AI solutions for different security needs. Initial partners include ZenDesk, Bell Canada, and AT&T, which are integrating Meta's AI-generated audio detection tools into their systems to identify scams, fraud, and phishing attempts.
Meta released CyberSecEval 4, an updated cybersecurity benchmark suite that includes CyberSOC Eval, developed with CrowdStrike to measure AI systems' efficacy in security operation centres, and AutoPatchBench, which evaluates AI systems' ability to automatically patch security vulnerabilities in native code.
The tools target enterprise security operations by providing automated threat detection capabilities and security classification systems. Meta's Automated Sensitive Doc Classification Tool automatically applies security labels to internal documents to prevent unauthorised access, and filter sensitive content from AI systems' RAG implementations.
The comprehensive security suite addresses growing enterprise concerns about AI system vulnerabilities as organisations increasingly deploy AI capabilities. Meta's approach of providing open source security tools positions the company to support enterprise AI adoption, while maintaining competitive advantages in AI safety and security infrastructure development.