As the debate around AI safety continues to heat up, Meta has gone public with a fresh batch of tools aimed at strengthening AI system security. The company’s latest update targets developers, cybersecurity experts, and even WhatsApp users. It brings together new open-source frameworks, security benchmarks, and a first look at private AI request processing. All of this came out during Meta’s ongoing push to make its Llama ecosystem safer, faster, and more transparent.

The highlight? A guardrail system called LlamaFirewall that promises real-time defenses against prompt injections, insecure code generation, and other risks that have emerged with large language models (LLMs) being used as autonomous agents.

LlamaFirewall brings real-time AI safety checks

Meta describes LlamaFirewall as a “final layer of defense” for AI systems. It comes with three guardrails: PromptGuard 2 for detecting jailbreaks and prompt injections, Agent Alignment Checks to audit the reasoning behind AI actions, and CodeShield, a static analysis tool to catch insecure code before it gets deployed.

What makes it interesting is the plug-and-play design. Developers can slot it into their own LLM-based apps and tweak it using custom scanners without digging deep into the internals. According to Meta, LlamaFirewall is already being used in production across its own AI workflows.

AutoPatchBench wants to fix software bugs with AI

Also part of Meta’s security push is AutoPatchBencha benchmark inside its CyberSecEval 4 suite. This is a tool that tests how well AI systems can automatically fix security bugs in code, especially vulnerabilities caught via fuzzing.

The benchmark uses a curated set of 136 C/C++ bugs from the ARVO dataset. It includes everything from buffer overflows to NULL pointer issues. AI systems are scored not just on whether they patch the crash but also on whether the fix introduces new problems.

It’s a major step for researchers and model builders working on autonomous software agents that can debug or repair real-world code. Meta says the idea is to move toward standardized testing for AI-driven bug fixing.

AI-generated voice detection and doc safety tools

Meta also launched tools like Llama Audio Watermark Detector and Llama Generated Audio Detector. These are meant to spot synthetic audio used in scams, phishing, or fraud. The company said it’s already working with Bell Canada, ZenDesk, and AT&T on this.

Another internal tool is the Automated Sensitive Doc Classification system, which labels documents to keep them out of AI training data or retrieval pipelines. It’s now available on GitHub.

Private Processing in WhatsApp gets early preview

Lastly, Meta showed off a glimpse of something called Private Processing. It’s a privacy-first feature that lets WhatsApp users run AI tools like message summarization without Meta being able to read the chats.

“We’re working with the security community to audit and improve our architecture,” the company said, promising to keep it open to researchers before full product launch.