Evidence · DiggingBeagle record
Heads of Security: Learned Attention Gates Improve Prompt Injection Robustness
Attention-head gating defense trained on channel-conditioned evidence and evaluated against adaptive prompt injection and benign performance. Reported robustness remains model- and benchmark-specific.
- Source role
- primary disclosure
Evidence record
Attention-head gating defense trained on channel-conditioned evidence and evaluated against adaptive prompt injection and benign performance. Reported robustness remains model- and benchmark-specific.
Claim-level citations (2)
- supportsLearned attention gates reduced selected prompt-injection successes with utility tradeoffs: The paper reports a large reduction in attack success on its RL-Hammer setting after adding learned channel-conditioned attention gates.
Abstract; Section 4 adaptive attacks
- supportsLearned attention gates reduced selected prompt-injection successes with utility tradeoffs: The authors separately evaluate benign utility and multiple agentic prompt-injection benchmarks instead of treating attack refusal as the only success criterion.
Section 4.2 and 4.3, utility and agentic benchmarks