A library for classifying strings as potential secrets.
virtualenv venv
source venv/bin/activate
pip install high-entropy-string
from high_entropy_string import PythonStringData
data = PythonStringData(
string='AKAI...',
node_type='assignment',
target='myvar',
patterns_to_ignore=[r'example.com'],
entropy_patterns_to_discount=[r'/BEGIN.*PUBLIC KEY/']
)
print(data.confidence)
print(data.severity)
This project is governed by Lyft's code of conduct. All contributors and participants agree to abide by its terms.
We require a CLA for code contributions, so before we can accept a pull request we need to have a signed CLA. Please visit our CLA service follow the instructions to sign the CLA.
The library classifies a string based on its liklihood of being a secret. We nudge the confidence and severity of the string based on criterea:
- Flags (ENTROPY_PATTERNS_TO_FLAG). Any Candidate that matches any regex in this list is automatically flagged as confidence/severity 3/3. If there's secret patterns you know conclusively are secrets, add them here.
- Discounts (ENTROPY_PATTERNS_TO_DISCOUNT). Any Candidate that matches a regex in this list is discounted. If the Candidate matches multiple regexes in this list, it may be discounted further. This discount is used in the confidence calculation.
- Secret hints (LOW_SECRET_HINTS, HIGH_SECRET_HINTS). If any target or caller matches a regex in these lists then it will be used as a hint that a Candidate is a secret. This hint is used in the confidence and severity calculations. LOW_SECRET_HINTS leads to a lower confidence increase and HIGH_SECRET_HINTS leads to a higher confidence increase.
- Safe functions (SAFE_FUNCTION_HINTS). Any Candidate that has a caller that matches any string in this list will will be discounted. This is used in the confidence calculation.
- Entropy. If a Candidate's confidence level can be more accurately gauged by a strings level of entropy, we calculate it and if the string has high entropy its confidence level is increased. This calculation is avoided if possible, as it's relatively expensive.
The concept is to eliminate noise while more easily identifying Candidates that may be secrets. Some help we'd love to have:
- Help with the discount regex list. The regexes in the list often match too much and there aren't enough that match common python strings.
- Help with the safe functions list (and the way we match the safe functions). There's a lot of python functions that rarely include secrets but often contain high entropy strings. We currently don't identify these function calls very well, which leads to higher noise.
- Add and improve string captures. We're not currently capturing all available strings in the AST and for some string captures we aren't capturing them as efficiently as we could. For instance with dicts, we capture info like: {'target': 'candidate'}, but don't capture: {'target': 'target': 'candidate'}, which could lead to better categorization.
Feel free to submit issues and pull requests for anything else you think would be useful as well.