SecPO cuts prompt injection attack success for agents
The point
Prompt injection lets hidden instructions in email, documents, and websites manipulate an agent. Meta-SecAlign, the prior best defended model, still saw attack success approaching 100 percent under adaptive attacks.
Why it matters
For builders, this offers training level defense for agents reading untrusted content.
Key details
- Adversarial training cut attack success 1.5 to 14.5 times versus Meta-SecAlign, SecPO cut it further 2.4 to 4.6 times versus DPO
- On AgentDojo, SecPO on Qwen3.6-27B hit 8.8 percent versus 78.8 percent for Meta-SecAlign
- No noticeable benign utility drop across nine benchmarks; code, data, and model released
Original ~8 min → 25 sec here Source: arXiv cs.AI, Nov 2, 2026