Heads of Security: Learned Attention Gates Improve Prompt Injection Robustness

Summary

Researchers introduce Heads of Security (HoS), a prompt injection defense that uses a small set of learned per-head attention gates to separate trusted instructions from untrusted data. Evaluated on Llama-3.3-70B-Instruct and GPT-OSS-20B, HoS cut the black-box RL-Hammer attack success rate to 10%, versus 97% on GPT-OSS-20B and 85% on SecAlign. It also preserved base-model utility on standard benchmarks, where SecAlign and ASIDE degraded in few-shot settings.

Topics

• Prompt injection and jailbreaks

Related items

• InfoThe ADLC maturity model: 5 stages of agentic development security | Endor LabsSimilar attack · Endor Labs Blog
• HighCVE-2026-118378: Tabularis SQLite driver SQL injection through MCP describe_table toolSimilar attack · NVD/CVE Database
• InfoThe Envelope Layer: An Underestimated Surface of LLM Agent DefenseSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• HighCVE-2026-117569: Kali-GPT OS command injection in execute_tool through MCP scanner parametersSimilar attack · NVD/CVE Database
• HighCVE-2026-117567: esp-mcp OS command injection through MCP tool parametersSimilar attack · NVD/CVE Database