Research map / Adversarial robustness and security / Attacks on text and graphs
Are aligned neural networks adversarially aligned?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| give NLP attack ARCA 10x more attacker-controlled tokens | ARCA · GPT-2 brute-force test cases | pass rate: 11.1% -> 30.6% | Helped |
| give NLP attack GBDA 10x more attacker-controlled tokens | GBDA · GPT-2 brute-force test cases | pass rate: 3.1% -> 9.5% | Helped |
| place adversarial tokens right before generation (nearby) instead of before the question (distant) | ARCA on LLaMA · Toxic output, 30 tokens | attack success rate: 2% -> 61% | Helped |
| alignment fine-tuning (Vicuna vs LLaMA) against NLP attack ARCA | LLaMA · Toxic output, nearby attack | attack success rate: 61% -> 8% | Hurt |
| use continuous adversarial image inputs instead of text-token attacks against Vicuna-based aligned models | aligned multimodal models · Toxic output | attack success rate: 8% -> 100% | Helped |
| switch target model architecture from Mini GPT-4/LLaMA Adapter to LLaVA | image attack · Toxic output | mean l2 distortion: 2.51 -> 0.86 | Helped |
| RLHF-tuned vs instruction-tuned Mini GPT-4 | Mini GPT-4 · Toxic output | mean l2 distortion: 2.51 -> 2.71 | No clear effect |
Closest papers by meaning
- Visual Adversarial Examples Jailbreak Aligned Large Language Models (AAAI 2024 · similarity 0.89)
- Fundamental Limitations of Alignment in Large Language Models (ICML 2024 · similarity 0.88)
- Semantic Representation Attack against Aligned Large Language Models (NeurIPS 2025 · similarity 0.88)
- Adversarial Attacks against Closed-Source MLLMs via Feature Optimal Alignment (NeurIPS 2025 · similarity 0.87)
- Scaling Laws for Adversarial Attacks on Language Model Activations and Tokens (ICLR 2025 · similarity 0.86)
- Safety Alignment Can Be Not Superficial With Explicit Safety Signals (ICML 2025 · 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.