Preprint · 2025
BountyBench: Dollar Impact of AI Agent Attackers and Defenders on Real-World Cybersecurity Systems
| Venue | arXiv |
|---|---|
| DOI | 10.48550/arXiv.2505.15216 |
| URL | http://arxiv.org/abs/2505.15216 |
| Topics | ReplicHAI / Agent Evals / 08 Safety & security |
| Tags | Computer Science - Artificial Intelligence, Computer Science - Computation and Language, Computer Science - Cryptography and Security, Computer Science - Machine Learning |
| archiveID | arXiv:2505.15216 |
| extra | arXiv:2505.15216 [cs.CR] |
| libraryCatalog | arXiv.org |
| shortTitle | BountyBench |
Abstract
AI agents have the potential to significantly alter the cybersecurity landscape. Here, we introduce the first framework to capture offensive and defensive cyber-capabilities in evolving real-world systems. Instantiating this framework with BountyBench, we set up 25 systems with complex, real-world codebases. To capture the vulnerability lifecycle, we define three task types: Detect (detecting a new vulnerability), Exploit (exploiting a given vulnerability), and Patch (patching a given vulnerability). For Detect, we construct a new success indicator, which is general across vulnerability types and provides localized evaluation. We manually set up the environment for each system, including installing packages, setting up server(s), and hydrating database(s). We add 40 bug bounties, which are vulnerabilities with monetary awards from \$10 to \$30,485, covering 9 of the OWASP Top 10 Risks. To modulate task difficulty, we devise a new strategy based on information to guide detection, interpolating from identifying a zero day to exploiting a given vulnerability. We evaluate 10 agents: Claude Code, OpenAI Codex CLI with o3-high and o4-mini, and custom agents with o3-high, GPT-4.1, Gemini 2.5 Pro Preview, Claude 3.7 Sonnet Thinking, Qwen3 235B A22B, Llama 4 Maverick, and DeepSeek-R1. Given up to three attempts, the top-performing agents are Codex CLI: o3-high (12.5% on Detect, mapping to \$3,720; 90% on Patch, mapping to \$14,152), Custom Agent: Claude 3.7 Sonnet Thinking (67.5% on Exploit), and Codex CLI: o4-mini (90% on Patch, mapping to \$14,422). Codex CLI: o3-high, Codex CLI: o4-mini, and Claude Code are more capable at defense, achieving higher Patch scores of 90%, 90%, and 87.5%, compared to Exploit scores of 47.5%, 32.5%, and 57.5% respectively; while the custom agents are relatively balanced between offense and defense, achieving Exploit scores of 17.5-67.5% and Patch scores of 25-60%.
BibTeX
@misc{zhang_bountybench_2025,
title = {{BountyBench}: {Dollar} {Impact} of {AI} {Agent} {Attackers} and {Defenders} on {Real}-{World} {Cybersecurity} {Systems}},
shorttitle = {{BountyBench}},
url = {http://arxiv.org/abs/2505.15216},
doi = {10.48550/arXiv.2505.15216},
abstract = {AI agents have the potential to significantly alter the cybersecurity landscape. Here, we introduce the first framework to capture offensive and defensive cyber-capabilities in evolving real-world systems. Instantiating this framework with BountyBench, we set up 25 systems with complex, real-world codebases. To capture the vulnerability lifecycle, we define three task types: Detect (detecting a new vulnerability), Exploit (exploiting a given vulnerability), and Patch (patching a given vulnerability). For Detect, we construct a new success indicator, which is general across vulnerability types and provides localized evaluation. We manually set up the environment for each system, including installing packages, setting up server(s), and hydrating database(s). We add 40 bug bounties, which are vulnerabilities with monetary awards from {\textbackslash}\$10 to {\textbackslash}\$30,485, covering 9 of the OWASP Top 10 Risks. To modulate task difficulty, we devise a new strategy based on information to guide detection, interpolating from identifying a zero day to exploiting a given vulnerability. We evaluate 10 agents: Claude Code, OpenAI Codex CLI with o3-high and o4-mini, and custom agents with o3-high, GPT-4.1, Gemini 2.5 Pro Preview, Claude 3.7 Sonnet Thinking, Qwen3 235B A22B, Llama 4 Maverick, and DeepSeek-R1. Given up to three attempts, the top-performing agents are Codex CLI: o3-high (12.5\% on Detect, mapping to {\textbackslash}\$3,720; 90\% on Patch, mapping to {\textbackslash}\$14,152), Custom Agent: Claude 3.7 Sonnet Thinking (67.5\% on Exploit), and Codex CLI: o4-mini (90\% on Patch, mapping to {\textbackslash}\$14,422). Codex CLI: o3-high, Codex CLI: o4-mini, and Claude Code are more capable at defense, achieving higher Patch scores of 90\%, 90\%, and 87.5\%, compared to Exploit scores of 47.5\%, 32.5\%, and 57.5\% respectively; while the custom agents are relatively balanced between offense and defense, achieving Exploit scores of 17.5-67.5\% and Patch scores of 25-60\%.},
urldate = {2026-11-06},
publisher = {arXiv},
author = {Zhang, Andy K. and Ji, Joey and Menders, Celeste and Dulepet, Riya and Qin, Thomas and Wang, Ron Y. and Wu, Junrong and Liao, Kyleen and Li, Jiliang and Hu, Jinghan and Hong, Sara and Demilew, Nardos and Murgai, Shivatmica and Tran, Jason and Kacheria, Nishka and Ho, Ethan and Liu, Denis and McLane, Lauren and Bruvik, Olivia and Han, Dai-Rong and Kim, Seungwoo and Vyas, Akhil and Chen, Cuiyuanxiu and Li, Ryan and Xu, Weiran and Ye, Jonathan Z. and Choudhary, Prerit and Bhatia, Siddharth M. and Sivashankar, Vikram and Bao, Yuxuan and Song, Dawn and Boneh, Dan and Ho, Daniel E. and Liang, Percy},
month = dec,
year = {2025},
note = {arXiv:2505.15216 [cs.CR]},
keywords = {Computer Science - Artificial Intelligence, Computer Science - Computation and Language, Computer Science - Cryptography and Security, Computer Science - Machine Learning},
}