Research map / Adversarial robustness and security / Black-box and transfer attacks
STACK: Adversarial Attacks on LLM Safeguard Pipelines
Ian R. McKenzie, Oskar John Hollinsworth, Tom Tseng, Xander Davies, Stephen Casper, Aaron David Tucker, Robert Kirk, Adam Gleave
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 |
|---|---|---|---|
| replace PAP with STACK staged attack | Gemma2 few-shot classifier pipeline · ClearHarm | ASR: 0% -> 71% | Improved |
| use transfer-STACK from a zero-shot Qwen3 proxy instead of direct access | Gemma2 few-shot classifier pipeline · ClearHarm | ASR: 71% -> 33% | Mixed |
| replace ShieldGemma with few-shot-prompted Gemma2 classifier | Qwen3-14B defense pipeline · ClearHarm | ASR: ASR reduced to 0% | Improved |
| evaluate PAP on ClearHarm instead of StrongREJECT | few-shot-prompted classifier pipelines · ClearHarm vs StrongREJECT | ASR: ASR plummets to zero | Improved |
| add STACK jailbreak strings to PAP on undefended model | Qwen3-14B undefended · ClearHarm | ASR: lower than pure PAP | Hurt |
| use ICJ as prefix rather than suffix in transfer template (ablation) | Transfer-STACK · ClearHarm | ASR | Improved |
Closest papers by meaning
- Claudini: Autoresearch Discovers State-of-the-Art Adversarial Attack Algorithms for LLMs (NeurIPS 2026 · similarity 0.86)
- AnchorRep: Defending LLMs Against Cross-Model Adversarial Transfer via Representation Repulsion (NeurIPS 2026 · similarity 0.85)
- Black-Box Adversarial Attacks on LLM-Based Code Completion (ICML 2025 · similarity 0.85)
- On Evaluating the Durability of Safeguards for Open-Weight LLMs (ICLR 2025 · similarity 0.85)
- Breaking Agent Backbones: Evaluating the Security of Backbone LLMs in AI Agents (ICLR 2026 · similarity 0.85)
- Soft Prompt Threats: Attacking Safety Alignment and Unlearning in Open-Source LLMs through the Embedding Space (NeurIPS 2024 · similarity 0.85)
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.
