LocalAlign: Near-target adversarial training improves the generalization of prompt injection defenses
Start here
Researchers from Nanyang Technological University and Wuhan University propose near-target adversarial training in the CCS 2026 paper LocalAlign. Targeting two types of insufficient generalization—optimizing injected text on the input side and approaching correct answers on the target side—they design a three-stage training process: warm-up, near-target adversarial example generation, and margin-aware alignment. On two models, across six out-of-distribution datasets, the ASR of optimization-free attacks all drop below 10%, while adaptive attacks remain more challenging.
Generated by AI from reports · Updated 7 days ago
Attention and developments
Attention, hourly
Now 38, 0% in 24 hours·Peak 42 (Nov 6 04:00)
Related research
Papers whose work the reports mention
- The reports mention Meta-SecAlignMeta-SecAlign: Training LLMs against Prompt Injection for Robust AgentsG.O.S.S.I.P Reading Recommendation 2026-11-05 LocalAlign, improving the generalization of prompt injection defenses through near-target adversarial training
arXiv · 2025-07-03 · Deep reading
- The reports mention SecAlignSecAlign: Defending Against Prompt Injection with Preference OptimizationG.O.S.S.I.P Reading Recommendation 2026-11-05 LocalAlign, improving the generalization of prompt injection defenses through near-target adversarial training
ACM CCS 2025 · 2024-10-07 · Deep reading
- The reports mention InjecAgentInjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated Large Language Model AgentsG.O.S.S.I.P Reading Recommendation 2026-11-05 LocalAlign, improving the generalization of prompt injection defenses through near-target adversarial training
arXiv · 2024-03-05 · Deep reading
Progress timeline
- WeChat · GoSSIPSelectedG.O.S.S.I.P Reading Recommendation 2026-11-05 LocalAlign, improving the generalization of prompt injection defenses through near-target adversarial training
Researchers from Nanyang Technological University and Wuhan University, in the CCS 2026 paper LocalAlign, propose improving the generalization of prompt injection defenses through near-target adversarial training. Addressing two types of insufficient generalization—input-side optimized injection text and target-side attack objectives close to correct answers—they design a three-stage training process: warm-up, near-target adversarial example generation, and margin-aware alignment. On Llama-3.1-8B-Instruct and Qwen3-4B-Instruct-2507, the ASR of optimization-free attacks on six out-of-distribution datasets all drops below 10%, while adaptive attacks remain more challenging; for Llama-3.1-8B-Instruct, the in-distribution adaptive GCG ASR on AlpacaFarm drops from 2.40% to 0.48%. Compared with Meta-SecAlign, normal capability drops by at most 2.00 percentage points; relative to the undefended original model, Qwen's MMLU-Pro drops from 56.74% to 49.35%. The paper and code are open-sourced.
Related discussion
- Older than 72 hours · Not included in the 72-hour figures
- WeChatDoonsecG.O.S.S.I.Pxa0 reading recommendationxa02026-11-05 LocalAlign: Improving the generalization of prompt injection defenses through near-target adversarial training(opens the original post in a new tab)
Attention trend
New sources per day
- Nov 5: 2 new sources (2 outlets, 0 accounts)
- Nov 6: 0 new sources (0 outlets, 0 accounts)
- Nov 7: 0 new sources (0 outlets, 0 accounts)
- Nov 8: 0 new sources (0 outlets, 0 accounts)
- Nov 9: 0 new sources (0 outlets, 0 accounts)
- Nov 10: 0 new sources (0 outlets, 0 accounts)
- Nov 11: 0 new sources (0 outlets, 0 accounts)
- Nov 12: 0 new sources (0 outlets, 0 accounts)
- Nov 13: 0 new sources (0 outlets, 0 accounts)