Meta releases Meta‑SecAlign-8B and Meta‑SecAlign-70B open‑weight models for prompt‑injection defense

Meta provided weights and tooling for model-level defenses against prompt injection, enabling researchers and practitioners to test agent security alongside utility.

In this brief: 2 sections 1 min read

- Two open-weight models released: Meta-SecAlign-8B and Meta-SecAlign-70B.
- Repository includes training, inference, evaluation tooling and benchmarks named AgentDojo and InjecAgent.
- Meta states the 70B model is comparable to certain closed models on its agentic security evaluations.
- SecAlign trains models to distinguish trusted instructions from untrusted input to avoid following injected commands.
- The report notes the 8B is based on Llama-3.1-8B-Instruct and the 70B on Llama-3.3-70B-Instruct.
- The repo claims the models are available for commercial use while some repository code is under a non-commercial license.

Read full analysis on letsdatascience.com

Useful?