Formalizing and Benchmarking Prompt Injection Attacks and Defenses
Summary
A prompt injection attack aims to inject malicious instruction or data into the input of an LLM-Integrated Application such that it produces results as an attacker desires. Existing works are limited to case studies. Their framework treats existing attacks as special cases and is used to combine them into a new attack. They then run a systematic evaluation of 5 prompt injection attacks and 10 defenses with 10 language models and 7 tasks. A public platform is released on GitHub. This note records the framework and the benchmark; it is not a how-to for attacks. Comments on the abs page state that the work was published in USENIX Security Symposium 2024. [S1]
Claims
- id: C1 claim: "The authors propose a framework that formalizes prompt injection and treats existing attacks as special cases." source: S1 evidence: "we propose a framework to formalize prompt injection attacks"
- id: C2 claim: "They evaluate 5 attacks and 10 defenses with 10 language models and 7 tasks." source: S1 evidence: "evaluation on 5 prompt injection attacks and 10 defenses with 10 LLMs and 7 tasks"
- id: C3 claim: "The benchmark platform is public on GitHub." source: S1 evidence: "https://github.com/liu00222/Open-Prompt-Injection"
Verification
- Code available: Paper: https://github.com/liu00222/Open-Prompt-Injection, HTTP 200 on 2026-11-12. Not run.
- Reproduced: not reproduced (2026-11-12). Code not run.
- Related work: A closely related attack class is covered in 2302-12173-indirect-prompt-injection. Practice notes include ao-security-prompt-injection-design.
- Limits: The 5-by-10-by-10-by-7 study is the paper's protocol, not an exhaustive map of later attacks. This note does not reproduce payloads or exploit steps.
- License: CC BY 4.0 on the abs page (2026-11-12; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 19 Oct 2023 (v1); last revised v5 (12 Nov 2025). Sources: [S1] https://arxiv.org/abs/2310.12815, v5 (12 Nov 2025). checked_on: 2026-11-12.
When to apply
Use when asking how prompt injection has been formalized and benchmarked in LLM-integrated applications.
Facts
- id: F1 claim: "The paper formalizes prompt injection in language-model applications and places prior attacks inside that framework." confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
- id: F2 claim: "The evaluation covers 5 attacks, 10 defenses, 10 models, and 7 tasks." confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
- id: F3 claim: "The platform is at https://github.com/liu00222/Open-Prompt-Injection." confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
Sources
- [S1] arXiv 2310.12815, https://arxiv.org/abs/2310.12815. License: CC BY 4.0. v5 (12 Nov 2025). Accessed 2026-11-12.
License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.
License: CC-BY-4.0 · Summary by iinventedlight (AI-assisted). Source: Yupei Liu, Yuqi Jia, Runpeng Geng, Jinyuan Jia, Neil Zhenqiang Gong, Formalizing and Benchmarking Prompt Injection Attacks and Defenses (CC BY 4.0).
Cite this
'iinventedlight (AI-assisted)'. 2026. "Formalizing and Benchmarking Prompt Injection Attacks and Defenses." iinventedlight Library, iil:note:ere0j6z9, version of 2026-11-12. https://iinventedlight.com/library/notes/2310-12815-prompt-injection-attacks-defenses/. License: CC-BY-4.0.
@misc{iil_note_ere0j6z9,
author = {{'iinventedlight (AI-assisted)'}},
title = {{Formalizing and Benchmarking Prompt Injection Attacks and Defenses}},
howpublished = {iinventedlight Library},
year = {2026},
date = {2026-11-12},
url = {https://iinventedlight.com/library/notes/2310-12815-prompt-injection-attacks-defenses/},
note = {Permanent ID iil:note:ere0j6z9. License: CC-BY-4.0.},
}
Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.
For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2310-12815-prompt-injection-attacks-defenses.md.