Research map / Adversarial robustness and security / Attacks on text and graphs
Security--Fidelity Tradeoffs: The Hidden Cost of Prompt Injection Defense
Mitchell Hermon, Rahul Gupta, Weitong Ruan, Ekraam Sabir, Haohan Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SecAlign defense | Llama 3.1 8B · SecFid | processing rate: 50.2 -> 67.0 | Improved |
| add DefensiveTokens defense | Llama 3.1 8B · SecFid | processing rate: 50.2 -> 37.7 | Hurt |
| add DefensiveTokens defense | Llama 3.1 8B · SecFid | execution rate: 49.1 -> 1.9 | Improved |
| add ASIDE defense | Qwen 2.5 7B · SecFid | processing rate: 54.7 -> 35.7 | Hurt |
| add SecAlign defense | Llama 3.3 70B · SecFid | ignore rate: 3.5 -> 29.0 | Hurt |
| add fidelity-aware DPO to SecAlign 8B | SecAlign 8B · SecFid core held-out | processing rate: 73.8 -> 83.0 | Improved |
| add fidelity-aware DPO to SecAlign 8B | SecAlign 8B · SecFid edit transfer (unseen task) | processing rate: 43.2 -> 80.6 | Improved |
| add fidelity-aware DPO to SecAlign 8B | SecAlign 8B · SecFid core held-out | execution rate: 0.0 -> 0.0 | No clear effect |
| scale Claude from Haiku to Opus | Claude 4.x · SecFid | security: 69.3 -> 81.8 -> 85.2 | Improved |
| scale Claude from Haiku to Opus | Claude 4.x · SecFid | fidelity: 77.3 -> 68.8 -> 67.7 | Hurt |
| scale Llama from 8B to 70B | Llama · SecFid | security: 50.9 -> 47.8 | Mixed |
| add reasoning (low/medium/high) | reasoning-capable models · SecFid | average security: +6.3, +9.1, +12.6 points | Improved |
Closest papers by meaning
- Assessing Vulnerabilities in State-of-the-Art Large Language Models Through Hex Injection (Student Abstract) (AAAI 2025 · similarity 0.87)
- ASPI: Seeking Ambiguity Clarification Amplifies Prompt Injection Vulnerability in LLM Agents (NeurIPS 2026 · similarity 0.86)
- AI Agents May Always Fall for Prompt Injections (NeurIPS 2026 · similarity 0.86)
- Bits Leaked per Query: Information-Theoretic Bounds for Adversarial Attacks on LLMs (NeurIPS 2025 · similarity 0.86)
- Expected Harm: Rethinking Safety Evaluation of (Mis)Aligned LLMs (NeurIPS 2026 · similarity 0.86)
- Skill-Inject: Measuring Agent Vulnerability to Skill File Attacks (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
