🧰 CyberSecEval2

The items that name it 3

• The state of AI for security: Measuring what matters most for building trust
  article2026-09-09It is written here as CyberSecEval 2.

  Meta benchmark measuring exploit generation

• meta-llama/PurpleLlama
  repo2026-07-14It is written here as CyberSec Eval 2.

  expands CyberSec Eval v1 by measuring an LLM's propensity to abuse a code interpreter, offensive cybersecurity capabilities, and susceptibility to prompt injection

• ReasAlign: Reasoning Enhanced Safety Alignment against Prompt Injection Attack
  paper

  Benchmark used to evaluate defenses against prompt-injected tasks in an open-ended setting