Heads of Security: Learned Attention Gates Improve Prompt Injection Robustness
arXiv cs.CL · 28 Oct 2026, 03:58 AEST
Heads of Security (HoS) is a new prompt injection defense that enhances instruction-data separation using learned attention gates. It demonstrates superior robustness against adaptive attacks compared to existing defenses like SecAlign++ and ASIDE.
Key facts
- HoS learns how each head should redistribute attention between trusted and untrusted tokens, enabling models to reason over untrusted content without following injected instructions.
- We show that HoS offers greater robustness to adaptive attacks, better performance on standard utility benchmarks, and matches top performance on standard agentic prompt injection benchmarks.
- The black-box RL-Hammer attack reaches 97% ASR against GPT-OSS-20B and 85% against SecAlign, but only 10% against HoS.
- HoS preserves base-model utility across standard multiple-choice benchmarks and prompting setups, whereas SecAlign and ASIDE degrade in few-shot settings.
- These results show that prompt injection robustness can be achieved and improved through targeted attention interventions.
Read aloud
Heads of Security HoS is a new prompt injection defense that enhances instruction-data separation using learned attention gates. It demonstrates superior robustness against adaptive attacks compared to existing defenses like SecAlign++ and ASIDE. HoS learns how each head should redistribute attention between trusted and untrusted tokens, enabling models to reason over untrusted content without following injected instructions. We show that HoS offers greater robustness to adaptive attacks, better performance on standard utility benchmarks, and matches top performance on standard agentic prompt injection benchmarks. The black-box RL-Hammer attack reaches ninety seven percent ASR against GPT-OSS-20B and eighty five percent against SecAlign, but only ten percent against HoS. HoS preserves base-model utility across standard multiple-choice benchmarks and prompting setups, whereas SecAlign and ASIDE degrade in few-shot settings. These results show that prompt injection robustness can be achieved and improved through targeted attention interventions.
Sources in this cluster
- Heads of Security: Learned Attention Gates Improve Prompt Injection Robustness
arXiv cs.CL · 28 Oct 2026, 03:58 AEST