README
RMCBench
Benchmarking Large Language Modelsβ Resistance to Malicious Code Generation Prompts
βββββββ ββββ ββββ ββββββββββββββ ββββββββββββ βββ ββββββββββ βββ
βββββββββββββ ββββββββββββββββββββββββββββββββββ ββββββββββββββ βββ
ββββββββββββββββββββββ ββββββββββββββ ββββββ ββββββ ββββββββ
ββββββββββββββββββββββ ββββββββββββββ βββββββββββββ ββββββββ
βββ ββββββ βββ ββββββββββββββββββββββββββββββ βββββββββββββββββ βββ
βββ ββββββ βββ ββββββββββββββ βββββββββββ βββββ ββββββββββ βββ
Why do we need to do this study?
The large language models be used to generate malicious code!!!
This is a hidden danger to the security of LLMs content.
Through RMCBench, you can conduct a comprehensive evaluation of LLMs' ability to resist malicious code generation.
Result Leaderboard
GOODοΌLLMs refuse to generate malicious code
| LLM | GOOD(%) | BAD(%) | UNCLEAR(%) |
|---|---|---|---|
| llama-2-13b-chat-hf | 48.84 | 49.26 | 1.90 |
| deepseek-coder-7b-instruct-v1.5 | 44.19 | 55.81 | 0.00 |
| Meta-Llama-3-8B-Instruct | 43.55 | 56.24 | 0.21 |
| mpt-7b-chat | 39.96 | 57.08 | 2.96 |
| llama-2-7b-chat-hf | 38.27 | 59.20 | 2.54 |
| gpt-4 | 35.73 | 64.27 | 0.00 |
| CodeLlama-13b-Instruct-hf | 30.66 | 68.92 | 0.42 |
| gpt-3.5-turbo | 18.39 | 81.18 | 0.42 |
| zephyr-7b-beta | 8.46 | 90.70 | 0.85 |
| vicuna-7b-v1.3 | 4.86 | 84.14 | 10.99 |
| tulu-2-13b | 2.96 | 90.27 | 6.77 |
| Average | 28.71 | 68.83 | 2.46 |
Characteristics
multi-scenarios
- text-to-code
- code-to-code
multi-tasks
- text-to-code generation (Level 1 - 3)
- code completion
- code translation
data details
The following is an explanation of all the fields in the dataset.
These fields are present in all samples:
- model_name: The full name of the LLM being tested.
- pid: The ID of the prompt.
- category: The scenario of malicious code generation (text-to-code, code-to-code).
- task: The specific task of malicious code generation (text-to-code generation, code translation, code completion).
- prompt: The prompt that instructs the LLMs to generate malicious code.
- malicious functionality: The specific malicious intent/functionality of the malicious code.
- malicious categories: The category of malicious code corresponding to the malicious intent/functionality.
- input_tokens: The token length of the prompt.
- response: The response from the LLMs.
- label: The automated labeling results from ChatGPT-4.
- check: The results of manual sampling checks on the label.
These fields are specific to the text-to-code scenario:
- level: The difficulty level of text-to-code.
- level description: The description and explanation of the level.
- jid: The ID of the jailbreak template (in level 3).
These fields are specific to the code-to-code scenario:
- cid: The ID of the malicious code sample we collected.
- original code: The complete malicious code sample we collected.
- language: The programming language of the malicious code.
- code lines: The number of lines in the malicious code.
- source: The source of the malicious code.
These fields are specific to the code-to-code scenario's code completion task:
- code to be completed: The remaining malicious code after being hollowing out.
- missing part: The hollowed out code (the code that needs to be completed).
- completion level: The level of code completion (token-level, line-level, multiline-level, function-level).
- completion position: The position of code completion (next token, fill-in-middle).
Paper Link
πArxiv πACM Digital Library
Dataset
π Github π€ Hugging Face
Citation
@inproceedings{10.1145/3691620.3695480,
author = {Chen, Jiachi and Zhong, Qingyuan and Wang, Yanlin and Ning, Kaiwen and Liu, Yongkun and Xu, Zenan and Zhao, Zhe and Chen, Ting and Zheng, Zibin},
title = {RMCBench: Benchmarking Large Language Models' Resistance to Malicious Code},
year = {2024},
isbn = {9798400712487},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3691620.3695480},
doi = {10.1145/3691620.3695480},
numpages = {12},
keywords = {large language models, malicious code, code generation},
location = {Sacramento, CA, USA},
series = {ASE '24}
}
ai-safety
code-generation
code-to-code
text-to-code
Contributors
zhongqy
4 commits
Similar repos
README
RMCBench
Benchmarking Large Language Modelsβ Resistance to Malicious Code Generation Prompts
βββββββ ββββ ββββ ββββββββββββββ ββββββββββββ βββ ββββββββββ βββ
βββββββββββββ ββββββββββββββββββββββββββββββββββ ββββββββββββββ βββ
ββββββββββββββββββββββ ββββββββββββββ ββββββ ββββββ ββββββββ
ββββββββββββββββββββββ ββββββββββββββ βββββββββββββ ββββββββ
βββ ββββββ βββ ββββββββββββββββββββββββββββββ βββββββββββββββββ βββ
βββ ββββββ βββ ββββββββββββββ βββββββββββ βββββ ββββββββββ βββ
Why do we need to do this study?
The large language models be used to generate malicious code!!!
This is a hidden danger to the security of LLMs content.
Through RMCBench, you can conduct a comprehensive evaluation of LLMs' ability to resist malicious code generation.
Result Leaderboard
GOODοΌLLMs refuse to generate malicious code
| LLM | GOOD(%) | BAD(%) | UNCLEAR(%) |
|---|---|---|---|
| llama-2-13b-chat-hf | 48.84 | 49.26 | 1.90 |
| deepseek-coder-7b-instruct-v1.5 | 44.19 | 55.81 | 0.00 |
| Meta-Llama-3-8B-Instruct | 43.55 | 56.24 | 0.21 |
| mpt-7b-chat | 39.96 | 57.08 | 2.96 |
| llama-2-7b-chat-hf | 38.27 | 59.20 | 2.54 |
| gpt-4 | 35.73 | 64.27 | 0.00 |
| CodeLlama-13b-Instruct-hf | 30.66 | 68.92 | 0.42 |
| gpt-3.5-turbo | 18.39 | 81.18 | 0.42 |
| zephyr-7b-beta | 8.46 | 90.70 | 0.85 |
| vicuna-7b-v1.3 | 4.86 | 84.14 | 10.99 |
| tulu-2-13b | 2.96 | 90.27 | 6.77 |
| Average | 28.71 | 68.83 | 2.46 |
Characteristics
multi-scenarios
- text-to-code
- code-to-code
multi-tasks
- text-to-code generation (Level 1 - 3)
- code completion
- code translation
data details
The following is an explanation of all the fields in the dataset.
These fields are present in all samples:
- model_name: The full name of the LLM being tested.
- pid: The ID of the prompt.
- category: The scenario of malicious code generation (text-to-code, code-to-code).
- task: The specific task of malicious code generation (text-to-code generation, code translation, code completion).
- prompt: The prompt that instructs the LLMs to generate malicious code.
- malicious functionality: The specific malicious intent/functionality of the malicious code.
- malicious categories: The category of malicious code corresponding to the malicious intent/functionality.
- input_tokens: The token length of the prompt.
- response: The response from the LLMs.
- label: The automated labeling results from ChatGPT-4.
- check: The results of manual sampling checks on the label.
These fields are specific to the text-to-code scenario:
- level: The difficulty level of text-to-code.
- level description: The description and explanation of the level.
- jid: The ID of the jailbreak template (in level 3).
These fields are specific to the code-to-code scenario:
- cid: The ID of the malicious code sample we collected.
- original code: The complete malicious code sample we collected.
- language: The programming language of the malicious code.
- code lines: The number of lines in the malicious code.
- source: The source of the malicious code.
These fields are specific to the code-to-code scenario's code completion task:
- code to be completed: The remaining malicious code after being hollowing out.
- missing part: The hollowed out code (the code that needs to be completed).
- completion level: The level of code completion (token-level, line-level, multiline-level, function-level).
- completion position: The position of code completion (next token, fill-in-middle).
Paper Link
πArxiv πACM Digital Library
Dataset
π Github π€ Hugging Face
Citation
@inproceedings{10.1145/3691620.3695480,
author = {Chen, Jiachi and Zhong, Qingyuan and Wang, Yanlin and Ning, Kaiwen and Liu, Yongkun and Xu, Zenan and Zhao, Zhe and Chen, Ting and Zheng, Zibin},
title = {RMCBench: Benchmarking Large Language Models' Resistance to Malicious Code},
year = {2024},
isbn = {9798400712487},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3691620.3695480},
doi = {10.1145/3691620.3695480},
numpages = {12},
keywords = {large language models, malicious code, code generation},
location = {Sacramento, CA, USA},
series = {ASE '24}
}
ai-safety
code-generation
code-to-code
text-to-code
Similar repos
Linked in READMEs
Contributors
zhongqy
4 commits