AI-Based Secrets Detection
Detect Secrets (API Tokens, Usernames, Passwords, etc.) Exposed on GitHub Repositories
Designed and Developed by Comcast Cybersecurity Research and Development Team

• Overview
• xGitGuard Workflow
• Features
  • Credential-Detection-Workflow
  • Keys/Token-Detection-Workflow
• Install
• Search Patterns
• Usage
  • Enterprise Github Secrets Detection
  • Public Github Secrets Detection
  • ML Model Training
• License

• Detecting Publicly Exposed Secrets on GitHub at Scale
  • xGitGuard is an AI-based system designed and developed by the Comcast Cybersecurity Research and Development team that detects secrets (e.g., API tokens, usernames, passwords, etc.) exposed on GitHub. xGitGuard uses advanced Natural Language Processing to detect secrets at scale and with appropriate velocity in GitHub repositories.
• What are Secrets?
  • Credentials
    • Usernames & passwords, server credentials, account credentials, etc.
  • Keys/Tokens
    • Service API tokens (AWS, Azure, etc), encryption keys, etc.

• Enterprise Credential Secrets Detection - Run Secret detection on the given GitHub Enterprise account

• Public Credential Secrets Detection - Run Secret detection on the GitHub Public account

  

• Enterprise Keys and Tokens Secrets Detection - Run Secret detection on the given GitHub Enterprise account

• Public Keys and Tokens Secrets Detection - Run Secret detection on the GitHub Public account

  

• Install Python >= v3.6

• Clone/Download the repository from GitHub

• Traverse into the cloned xGitGuard folder

  cd xGitGuard
  

• Install Python Dependency Packages

  python -m pip install -r requirements.txt
  

• Check for Outdated Packages

  pip list --outdated
  

• There are two ways to define configurations in xGitGuard

  • Config Files
  • Command Line Inputs

• For Enterprise Github Detection (Secondary Keyword + Extension) under config directory

  • Secondary Keyword: secondary_keys.csv file or User Feed - list of Keys & Tokens
  • Secondary Keyword: secondary_creds.csv file or User Feed - list of Credentials
  • Extension: extensions.csv file or User Feed - List of file Extensions

• For Public Github Detection (Primary Keyword + Secondary Keyword + Extension) under config directory

  • Primary Keyword: primary_keywords.csv file or User Feed - list of primary Keys
  • Secondary Keyword: secondary_keys.csv file or User Feed - list of Keys & Toekns
  • Secondary Keyword: secondary_creds.csv file or User Feed - list of Credentials
  • Extension: extensions.csv file or User Feed - List of file Extensions

• Enterprise Github Secrets Detection
  • Enterprise Credential Secrets Detection
  • Enterprise Keys and Tokens Secrets Detection
• Public Github Secrets Detection
  • Public Credential Secrets Detection
  • Public Keys and Tokens Secrets Detection

• Setup the system Environment variable below for accessing GitHub
  • GITHUB_ENTERPRISE_TOKEN - Enterprise GitHub API Token with full scopes of repository and user.
    • Refer to the GitHub documentation How To Get GitHub API Token for help
• Update the following configs with your Enterprise Name in config file xgg_configs.yaml in config Data folder xgitguard\config\*
  • enterprise_api_url: https://github.<<Enterprise_Name>>.com/api/v3/search/code
  • enterprise_pre_url: https://github.<<Enterprise_Name>>.com/api/v3/repos/
  • url_validator: https://github.<<Enterprise_Name>>.com/api/v3/search/code
  • enterprise_commits_url: https://github.<<Enterprise_Name>>.com/api/v3/repos/{user_name}/{repo_name}/commits?path={file_path}

• Traverse into the github-enterprise script folder

  cd github-enterprise
  

By default, the Credential Secrets Detection script runs for given Secondary Keywords and extensions without ML Filter.

# Run with Default configs
python enterprise_cred_detections.py

xGitGuard also has an additional ML filter where users can collect their organization/targeted data and train their model. Having this ML filter helps to reduce the false positives from the detection.

User Needs to follow the below process to collect data and train the model to use ML filter.

• Follow ML Model Training

NOTE :

• To use ML Filter, ML training is mandatory. This includes data collection, feature engineering & model persisting.
• This process is going to be based on user requirements. It can be one time or if the user needs to improve the data, then needs to be done periodically.

# Run for given Secondary Keyword and extension with ML model,
python enterprise_cred_detections.py -m Yes

Run usage:
enterprise_cred_detections.py [-h] [-s Secondary Keywords] [-e Extensions] [-m Ml prediction][-u Unmask Secret][-l Logger Level] [-c Console Logging]

optional arguments:
  -h, --help            show this help message and exit
  -s Secondary Keywords, --secondary_keywords Secondary Keywords
                          Pass the Secondary Keywords list as a comma-separated string
  -e Extensions, --extensions Extensions
                          Pass the Extensions list as a comma-separated string
  -m ML Prediction, --ml_prediction ML Prediction
                          Pass the ML Filter as Yes or No. Default is No
  -u Set Unmask, --unmask_secret To write secret unmasked, then set Yes
                          Pass the flag as Yes or No. Default is No
  -l Logger Level, --log_level Logger Level
                          Pass the Logging level as for CRITICAL - 50, ERROR - 40 WARNING - 30 INFO - 20 DEBUG - 10. Default is 20
  -c Console Logging, --console_logging Console Logging
                          Pass the Console Logging as Yes or No. Default is Yes

• Inputs used for search and scan

  Note: Command-line argument keywords have precedence over config files (Default). If no keywords are passed in cli, data from config files will be used for the search.

  • secondary_creds.csv file has a default list of credential relevant patterns for search, which can be updated by users based on their requirement.
  • extensions.csv file has a default list of file extensions to be searched, which can be updated by the users based on their requirement.

• GitHub search pattern for above examples: password +extension:py

By default, the Keys and Tokens Secrets Detection script runs for given Secondary Keywords and the extensions without ML Filter.

# Run with Default configs
python enterprise_key_detections.py

xGitGuard also has an additional ML filter where users can collect their organization/targeted data and train their model. Having this ML filter helps in reducing the false positives from the detection.

The user needs to follow the below process to collect data and train the model to use ML filter.

• Follow ML Model Training

NOTE :

• To use ML filter, ML training is mandatory. It includes data collection, feature engineering & model persisting.
• This process is going to be based on user requirements. It can be one time or if the user needs to improve the data, then it needs to be done periodically.

# Run for given Secondary Keyword and extension with ML model
python enterprise_key_detections.py -m Yes

Run usage:
enterprise_key_detections.py [-h] [-s Secondary Keywords] [-e Extensions] [-m Ml prediction][-u Unmask Secret][-l Logger Level] [-c Console Logging]

optional arguments:
  -h, --help            show this help message and exit
  -s Secondary Keywords, --secondary_keywords Secondary Keywords
                          Pass the Secondary Keywords list as a comma-separated string
  -e Extensions, --extensions Extensions
                          Pass the Extensions list as a comma-separated string
  -m ML Prediction, --ml_prediction ML Prediction
                          Pass the ML Filter as Yes or No. Default is No
  -u Set Unmask, --unmask_secret To write secret unmasked, then set Yes
                          Pass the flag as Yes or No. Default is No
  -l Logger Level, --log_level Logger Level
                          Pass the Logging level as for CRITICAL - 50, ERROR - 40 WARNING - 30 INFO - 20 DEBUG - 10. Default is 20
  -c Console Logging, --console_logging Console Logging
                          Pass the Console Logging as Yes or No. Default is Yes

• Inputs used for search and scan

  Note: Command-line argument keywords have precedence over config files (Default). If no keywords are passed in cli, data from the config files will be used for search.

  • secondary_keys.csv file will have a default list of key relevant patterns for search, which can be updated by the users based on their requirement.
  • extensions.csv file has a default list of file extensions to be searched, which can be updated by the users based on their requirement.

• GitHub search pattern for above examples: api_key +extension:py

• Credentials

    1. Hashed Url Files: xgitguard\output\*_enterprise_hashed_url_creds.csv
        - List previously Processed Search urls. Urls stored will be skipped in next run to avoid re processing.
    2. Secrets Detected: xgitguard\output\*_xgg_enterprise_creds_detected.csv
    3. Log File: xgitguard\logs\enterprise_key_detections_*yyyymmdd_hhmmss*.log
  

• Keys & Tokens

    1. Hashed Url Files: xgitguard\output\*_enterprise_hashed_url_keys.csv
        - List previously Processed Search urls. Urls stored will be skipped in next run to avoid re processing.
    2. Secrets Detected: xgitguard\output\*_xgg_enterprise_keys_detected.csv
    3. Log File: xgitguard\logs\enterprise_key_detections_*yyyymmdd_hhmmss*.log
  

• Setup the Environment variable below for accessing GitHub
  • GITHUB_TOKEN - Public GitHub API Token with full scopes of the repository and user.
    • Refer to GitHub Docs How To Get GitHub API Token for help
• Config data folder xgitguard\config\*

• Traverse into the github-public script folder

  cd github-public
  

Note: User needs to remove the sample content from primary_keywords.csv and add primary keywords like targeted domain names to be searched in public GitHub.

By default, Credential Secrets Detection script runs for given Primary Keyword, Secondary Keyword, and extension without ML Filter.

# Run with Default configs
python public_cred_detections.py

xGitGuard also has an additional ML filter, where users can collect their organization/targeted data and train their model. Having this ML filter helps in reducing the false positives from the detection.

The user needs to follow the below process to collect data and train the model to use ML filter.

• Follow ML Model Training

NOTE :

• To use ML Feature, ML training is mandatory. It includes data collection, feature engineering & model persisting.

# Run for given Primary Keyword, Secondary Keyword, and extension with ML model
python public_cred_detections.py -m Yes

Run usage:
usage: public_cred_detections.py [-h] [-p Primary Keywords] [-s Secondary Keywords] [-e Extensions] [-m Ml prediction][-u Unmask Secret] [-l Logger Level] [-c Console Logging]

optional arguments:
-h, --help show this help message and exit
-p Primary Keywords, --primary_keywords Primary Keywords
Pass the Primary Keywords list as a comma-separated string
-s Secondary Keywords, --secondary_keywords Secondary Keywords
Pass the Secondary Keywords list as a comma-separated string
-e Extensions, --extensions Extensions
Pass the Extensions list as a comma-separated string
-m ML Prediction, --ml_prediction ML Prediction
                          Pass the ML Filter as Yes or No. Default is No
-u Set Unmask, --unmask_secret To write secret unmasked, then set Yes
                          Pass the flag as Yes or No. Default is No
-l Logger Level, --log_level Logger Level
Pass the Logging level as for CRITICAL - 50, ERROR - 40 WARNING - 30 INFO - 20 DEBUG - 10. Default is 20
-c Console Logging, --console_logging Console Logging
Pass the Console Logging as Yes or No. Default is Yes

• Inputs used for search and scan

  Note: Command line argument keywords have precedence over config files (Default). If no keywords are passed in cli, config files data will be used for search.

  • primary_keywords.csv file will have a default list of primary keyword-relevant patterns for search
  • secondary_creds.csv file will have a default list of credential relevant patterns for search, which can be updated by the users based on their requirement.
  • extensions.csv file has a default list of file extensions to be searched, which can be updated by the users based on their requirement.

• GitHub search pattern for above examples: abc.xyz.com password +extension:py

By default, Keys and Tokens Secret Detection script runs for given Primary Keyword, Secondary Keyword and extension without ML Filter.

# Run with Default configs
python public_key_detections.py

xGitGuard also has an additional ML filter, where users can collect their organization/targeted data and train their model. Having this ML filter helps in reducing the false positives from the detection.

The user needs to follow the below process to collect data and train the model to use ML filter.

• Follow ML Model Training

NOTE : To use ML Feature, ML training is mandatory. It includes data collection,feature engineering & model persisting.

# Run for given  Primary Keyword, Secondary Keyword, and extension with ML model,
python public_key_detections.py -m Yes

usage:
public_key_detections.py [-h] [-s Secondary Keywords] [-e Extensions] [-m Ml prediction][-u Unmask Secret] [-l Logger Level] [-c Console Logging]

optional arguments:
-h, --help show this help message and exit
-s Secondary Keywords, --secondary_keywords Secondary Keywords
Pass the Secondary Keywords list as a comma-separated string
-e Extensions, --extensions Extensions
Pass the Extensions list as a comma-separated string
-m ML Prediction, --ml_prediction ML Prediction
                          Pass the ML Filter as Yes or No. Default is No
-u Set Unmask, --unmask_secret To write secret unmasked, then set Yes
                          Pass the flag as Yes or No. Default is No
-l Logger Level, --log_level Logger Level
Pass the Logging level as for CRITICAL - 50, ERROR - 40 WARNING - 30 INFO - 20 DEBUG - 10. Default is 20
-c Console Logging, --console_logging Console Logging
Pass the Console Logging as Yes or No. Default is Yes

• Inputs used for search and scan

  Note: Command line argument keywords have precedence over config files (Default). If no keywords are passed in cli, config files data will be used for search.

  • primary_keywords.csv file will have a default list of primary keyword-relevant patterns for search, which can be updated by the users based on their requirement.
  • secondary_keys.csv file will have a default list of tokens & keys relevant patterns for search, which can be updated by the users based on their requirement.
  • extensions.csv file has a default list of file extensions to be searched, which can be updated by the users based on their requirement.

• GitHub search pattern for above examples: abc.xyz.com api_key +extension:py

• Credentials

    1. Hashed Url Files: xgitguard\output\*_public_hashed_url_creds.csv
        - List pf previously Processed Search urls. Urls stored will be skipped in next run to avoid re processing.
    2. Secrets Detected: xgitguard\output\*_xgg_public_creds_detected.csv
    3. Log File: xgitguard\logs\public_key_detections_*yyyymmdd_hhmmss*.log
  

• Keys & Tokens

    1. Hashed Url Files: xgitguard\output\*_public_hashed_url_keys.csv
        - List pf previously Processed Search urls. Urls stored will be skipped in next run to avoid re processing.
    2. Secrets Detected: xgitguard\output\*_xgg_public_keys_detected.csv
    3. Log File: xgitguard\logs\public_key_detections_*yyyymmdd_hhmmss*.log
  

Note: By Default, the detected secrets will be masked to hide sensitive data. If needed, user can skip the masking to write raw secret using command line argument -u Yes or --unmask_secret Yes. Refer command line options for more details.

To use ML Feature, ML training is mandatory. It includes data collection, feature engineering & model persisting.

Note: Labelling the collected secret is an important process to improve the ML prediction.

• Traverse into the "ml_training" folder

  cd ml_training
  

Traverse into the "data collector" folder under ml_training

  cd ml_data_collector\github-enterprise-ml-data-collector

• Credentials

  1. Run for given Secondary Keywords and extensions

     python enterprise_cred_data_collector.py
     

  2. To run with other parameters, please use help.

     python enterprise_cred_data_collector.py  -h
     

  3. Training data for Enterprise Creds collected will be placed in xgitguard\output\cred_train_source.csv folder

• Keys & Tokens

  1. Run for given Secondary Keywords and extensions,

     python enterprise_key_data_collector.py
     

  2. To run with other parameters, please use help.

     python enterprise_key_data_collector.py  -h
     

  3. Training data for Enterprise Keys and Tokens collected will be placed in xgitguard\output\key_train_source.csv folder

1. By default all the data collected will be labeled as 1 under column "Label" in collected training data indicating the collected secret as a valid one.
2. User needs to review each row in the collected data and update the label value. i.e: if the user thinks collected data is not a secret, then change the value to 0 for that particular row.
3. By doing this, ML will have quality data for the model to reduce false positives.

Traverse into the "ml_training" folder

• Credentials

  1. Run with option cred for engineering collected cred data

     python ml_feature_engineering.py cred
     

  2. By default in Enterprise mode, input will be cred_train_source.csv
  3. Engineered data for Enterprise Creds output will be placed in xgitguard\output\cred_train.csv folder

• Keys & Tokens

  1. Run with option cred for engineering collected keys & tokens data

     python ml_feature_engineering.py key
     

  2. By default in Enterprise mode, input will be key_train_source.csv
  3. Engineered data for Enterprise Keys & Tokens output will be placed in xgitguard\output\key_train.csv folder

Traverse into the "ml_training" folder

• Run training with Cred Training Data and persist model

  python model.py cred
  

• Run training with Key Training Data and persist model

  python model.py key
  

• For help on command line arguments, run

  python model.py  -h
  

  Note: If persisted model xgitguard\output\xgg_*.pickle is not present in the output folder, then use engineered data to create a model and persist it.

To use ML Feature, ML training is mandatory. It includes data collection, feature engineering & model persisting.

Note: Labelling the collected secret is an important process to use the ML effectively.

• Traverse into the "models" folder

  cd ml_training
  

Traverse into the "data collector" folder

cd ml_training\ml_data_collector\github-public-ml-data-collector

Note: User needs to remove the sample content from primary_keywords.csv and add primary keywords like targeted domain names to be searched in public GitHub.

• Credentials

  1. Run for given Primary Keywords, Secondary Keywords, and extensions

     python public_cred_data_collector.py
     

  2. To run with other parameters, please use help.

     python public_cred_data_collector.py -h
     

  3. Training data for Public Creds collected will be placed in xgitguard\output\public_cred_train_source.csv folder

• Keys & Tokens

  1. Run for given Primary Keywords, Secondary keywords, and extensions

     python public_key_data_collector.py
     

  2. To run with other parameters, please use help.

     python public_key_data_collector.py  -h
     

  3. Training data for Public Keys and Tokens collected will be placed in xgitguard\output\public_key_train_source.csv folder

Note: The data collection for public GitHub is optional.

• If targeted data collected from Enterprise is enough to use, then we can skip the data collection & Label review process

1. By default, all the data collected will be labeled as 1 under column "Label" in collected training data indicating the collected secret as a valid one.
2. User needs to review each row in the collected data and update the label value. i.e: if the user thinks collected data is not a secret, then change the value to 0 for that particular row.
3. By doing this, ML will have quality data for the model to reduce false positives.

Note: Labelling the collected secret is an important process to use the ML effectively.

Traverse into the "ml_training" folder

• Credentials

  1. Run with option cred for engineering collected cred data with public source data.

     python ml_feature_engineering.py cred -s public
     

  2. In public mode, input will be public_cred_train_source.csv
  3. Engineered data for Public Creds output will be placed in xgitguard\output\public_cred_train.csv folder

• Keys & Tokens

  1. Run with option cred for engineering collected keys & tokens data with public source data.

     python ml_feature_engineering.py key -s public
     

  2. In public mode, input will be public_key_train_source.csv
  3. Engineered data for Public Keys & Tokens output will be placed in xgitguard\output\public_key_train.csv folder

Note:

• Data collection & feature engineering for public GitHub scan is optional.
• When public training data not available, feature engineering will use enterprise source data.

Traverse into the "ml_training" folder

• Run training with Cred Training Data and persist model with public source data

  python model.py cred -s public
  

• Run training with Key Training Data and persist model with public source data

  python model.py key -s public
  

• For help on command line arguments, run

  python model.py  -h
  

  Note:

  • If persisted model xgitguard\output\public_*xgg*.pickle is not present in the output folder, then use feature engineered data to create a model and persist it.
  • By default, when feature engineered data collected in Public mode not available, then model creation will be using enterprise-based engineered data.

• Users can update confidence_values.csv based on secondary_keys, secondary_creds, extensions value and give scoring from level 0 (lowest) to 5 (highest) to denote associated keyword suspiciousness.
• If users need to add any custom/new secondary creds/keys or extensions to the config files, then the same has to be added in the confidence_values.csv file with respective score level.
• Stop Words provided in config files are very limited and generic.Users need to update stop_words.csv with keywords considered has false postives to filter it out from the detections.
• Users can add additional extensions to extensions.csv to search types of files other than the default list.
• Users can enhance secondary_creds.csv/secondary_keys.csv by adding new patterns to do searches other than the default list.
• Users need to add primary keywords for public search in primary_keywords.csv after removing the sample content.
• In case of GitHub API calls resulting in 403 due to API rate-limit, increase the throttle timeout (github.throttle_time: 10) in the config ("config/xgg_configs.yaml)".

Licensed under the Apache 2.0 license.