As AI systems evolve, the challenges of ensuring safety grow alongside them. Zacharie Delpierre Coudert and Spencer Whitman from Meta’s Llama Trust & Safety team recently discussed on developing AI models with built-in safeguards. With the release of Llama 3.1, Meta is advancing its approach to AI safety, focusing on system-level protections that developers can use to build secure applications from the ground up.

According to Delpierre Coudert: “It’s exciting to see LLMs (Large Language Models) accomplish more complex tasks, but this evolution also brings new safety and security challenges.” He noted the shift from simple chatbot interactions to AI agents capable of executing tasks, which opens up new vulnerabilities. “We’ve evolved our safety tools with this shift,” he added, highlighting Meta’s commitment to addressing these risks.

One of the key tools Meta developed is Llama Guard, a content moderation system designed to filter unsafe inputs and outputs.

“Llama Guard has been upgraded to support new features like tool calls and multilingual capabilities,” Delpierre Coudert explained. The team’s approach includes more flexibility for developers, allowing them to adapt these safeguards to specific use cases.

Whitman stressed the importance of modularizing AI safety: “You can’t apply the same safety measures for every use case, so we’ve created tools like Prompt Guard to detect prompt injections or jailbreak attempts.” This allows developers to tailor safety mechanisms for their unique applications. “Prompt Guard is fast, lightweight, and helps ensure that AI systems aren’t exploited through subtle, harmful inputs,” he said.

Beyond content moderation, Meta’s Code Shield is another critical layer, ensuring secure code generation from AI models.

“Code Shield helps filter out insecure coding practices, making sure AI-generated code is safe,” Whitman added.

With these advancements, Meta is not only fostering innovation but also giving developers the tools to build AI responsibly.

“We want developers to have control over the safety of their applications,” Whitman concluded. “Our mission is to provide the flexibility and resources needed to create secure, innovative systems that can be trusted.”