Meta

Sources: Meta (homepage) · Meta AI

Meta maintains the Purple Llama project — the wiki’s main point of contact with Meta’s AI-security work. Purple Llama bundles the LlamaFirewall agent-guardrail framework and the CyberSecEval offensive-capability benchmark suite under one purple-team banner (red-team measurement plus blue-team defense).

Joshua Saxe tech-led AI security at Meta for roughly four years — including the prompt-injection-defense work that became LlamaFirewall, for which he is the corresponding author — and presented Measuring Security-Agent Effectiveness at Unprompted March 2026.