Llama Guard 3, Prompt Guard and CodeShield overview

Uploaded on August 9, 2024, this discussion covers Meta's safety tooling around the Llama 3.1 release, including Llama Guard 3 and CyberSecEval 3. Meta's trust and safety team explains why an AI agent that reads external data and calls tools needs safeguards beyond the model's built-in safety training. The discussion concerns that release period, rather than current capabilities or installation steps.

The speakers describe Llama Guard 3 as an input and output content moderation model with support for eight languages, a longer context and checks for some tool calls, including search and code interpreter use. Prompt Guard addresses a different problem: attempts to redirect a model through user prompts or external text. They describe an 86 million parameter multilingual classifier that labels text as benign, prompt injection or jailbreak. One speaker says it reads 512 tokens at a time, with some uncertainty about that figure.

For a coding assistant, CodeShield checks generated code for known insecure patterns. The team says the release supported seven programming languages and about 50 common weakness enumerations. It uses static analysis rather than another language model; applications can suppress flagged suggestions or request replacement output.

The discussion also points developers to Purple Llama, deployment notebooks in Llama recipes and CyberSecEval 3 for evaluating cybersecurity behavior. The speakers encourage customization and testing for each application's needs; their safety claims do not establish that these tools prevent every attack.