Purple Llama (CyberSecEval 1-4, AutoPatchBench, CyberSOCEval, LlamaFirewall, Llama Guard)

Benchmark / dataset

Meta's security suite: insecure-code, cyberattack helpfulness, prompt injection, autonomous ops, patching, SOC tasks; plus guardrails.

Category: Cyber defense & threat intel
Insurance exposure: Breach/cyber
Publisher: Meta
Publisher type: Big tech
License: unknown
Status: Active (<90d)
Last commit: 2026-10-12
Stars: 4433
Notes: Broadest single cyber eval suite from a lab. CyberSOCEval co-built with CrowdStrike.

Wanich Risk Lab, 2026.