🧰 Meta-SecAlign

The items that name it 3

• Awesome Offensive AI Agentic Landscape
  repo2026-08-01

  Group B security-specialized model aligned against prompt injection/defense

• ReasAlign: Reasoning Enhanced Safety Alignment against Prompt Injection Attack
  paperIt is written here as Meta SecAlign.

  Prior state-of-the-art guardrail/defensive model against prompt injection attacks, used as the comparison baseline (achieved 56.4% utility and 74.4% ASR on CyberSecEval2 vs ReasAlign's 94.6% utility and 3.6% ASR)

• PISmith: Reinforcement Learning-based Red Teaming for Prompt Injection Defenses
  paperIt is written here as Meta-SecAlign (Chen et al., 2025c; DPO-fine-tuned LLM defense).

  Chen et al., 2025c; DPO-fine-tuned LLM defense