This document describes experimental features and how to try them. Have fun, file bugs, tweak the code, and most importantly share your thoughts!


Hands down the best way to enforce a code standard is to just automatically fix it. Semgrep's rule format supports a fix: key that supports metavariable replacement, much like message fields. This allows for value capture and rewriting.

The autofix can be applied directly to the file using the --autofix flag, or you can use both the --autofix and --dryrun flags to test the autofix.

Example autofix (see in editor here):

- id: use-sys-exit
  - python
  message: |
    Use `sys.exit` over the python shell `exit` built-in. `exit` is a helper
    for the interactive shell and is not be available on all Python implementations.
  pattern: exit($X)
  fix: sys.exit($X)
  severity: WARNING

Autofix with regular expression replacement

A variant on the experimental fix key is fix-regex, which applies regular expression replacements (think sed) to matches found by Semgrep.

fix-regex has two required fields:

  • regex specifies the regular expression to replace within the match found by Semgrep
  • replacement specifies what to replace the regular expression with.

fix-regex also takes an optional count field, which specifies how many occurrences of regex to replace with replacement, from left-to-right and top-to-bottom. By default, fix-regex will replace all occurrences of regex. If regex does not match anything, no replacements are made.

The replacement behavior is identical to the re.sub function in Python. See these Python docs for more information.

An example rule with fix-regex is shown below. regex uses a capture group to greedily capture everything up to the final parenthesis in the match found by Semgrep. replacement replaces this with everything in the capture group (\1), a comma, timeout=30, and a closing parenthesis. Effectively, this adds timeout=30 to the end of every match.

- id: python.requests.best-practice.use-timeout.use-timeout
  - pattern-not: requests.$W(..., timeout=$N, ...)
  - pattern-not: requests.$W(..., **$KWARGS)
  - pattern-either:
    - pattern: requests.request(...)
    - pattern: requests.get(...)
    - pattern: requests.post(...)
    - pattern: requests.put(...)
    - pattern: requests.delete(...)
    - pattern: requests.head(...)
    - pattern: requests.patch(...)
    regex: '(.*)\)'
    replacement: '\1, timeout=30)'
  message: |
    'requests' calls default to waiting until the connection is closed.
    This means a 'requests' call without a timeout will hang the program
    if a response is never received. Consider setting a timeout for all
  languages: [python]
  severity: WARNING

Apply Semgrep autofix direclty to a file


Equivalences allows equivalent code patterns to be defined (i.e. a commutative property: $X + $Y <==> $Y + $X).

Use the equivalences top-level key and one - equivalence: key for each equivalence.

For example (see in editor here):

  - id: open-redirect
    languages: [python]
      - equivalence: request.$W.get(...) ==> request.$W(...)
      - pattern-inside: |
          def $FUNC(...):
      - pattern-not-inside: |
          def $FUNC(...):
      - pattern-not-inside: |
          if <... django.utils.http.is_safe_url(...) ...>:
      - pattern-either:
        - pattern: django.shortcuts.redirect(..., request.$W.get(...), ...)
        - pattern: django.shortcuts.redirect(..., $S.format(..., request.$W.get(...), ...), ...)
        - pattern: django.shortcuts.redirect(..., $S % request.$W.get(...), ...)
        - pattern: django.shortcuts.redirect(..., f"...{request.$W.get(...)}...", ...)
    message: "Open redirect detected."
    severity: WARNING

Data-flow analysis

Semgrep can perform intra-procedural flow-sensitive analyses. The data-flow engine still has several limitations, therefore expect both false positives and false negatives. False positives could be removed by using pattern-not.

A non-exhaustive list of current limitations:

  • The analyses are not aware of aliasing.
  • The analyses do not track individual elements in data structures, although there is limited support for record fields.
  • break, continue, and switch statements are not properly handled yet.
  • try-catch-finally is only partially supported, not all possible execution paths are considered.

As of now, data-flow analysis is used for taint tracking and constant propagation.

Taint tracking

The Python CLI has support for within (intra) file taint tracking. A taint-tracking rule uses the mode: taint key-value pair and replaces the typical top-level pattern keys with pattern-sources and pattern-sinks (required) and pattern-sanitizers (optional). For example:

- id: rule_id
  mode: taint
    - source(...)
    - source1(...)
    - sink(...)
    - sink1(...)
    - eval(...)
    - sanitize(...)
    - sanitize1(...)
  message: A user input source() went into a dangerous sink()
  languages: [python]
  severity: WARNING

A file containing the rule shown above can be found in the Semgrep repo. To see this taint-tracking example in action, run the above Semgrep rule on the taint test. Assuming you are in the Semgrep project’s root working directory:

semgrep --config https://raw.githubusercontent.com/returntocorp/semgrep/develop/semgrep-core/data/basic_tainting.yml ./semgrep-core/tests/OTHER/TAINTING/tainting.py

Constant propagation

Constant propagation is intra-procedural and tracks whether a variable must carry a constant value at each point in the program.

For example, we can find calls to eval on arbitrary strings in Python as follows:

- id: eval_arbitrary
    - pattern-either:
      - pattern: |
    - pattern-not: |
  message: |
    eval() on arbitrary non-constant string
  severity: WARNING

In the following code, Semgrep will only warn about eval(x). Variable x is not a constant because it may take the value of the input arg. Variable y is known to be a constant despite we do not know its exact value, it may be either "1" or "2". Pattern eval("...") matches eval(y), but neither pattern eval("1") nor eval("2") will match eval(y). Finally, variable z is a constant and it is known to be "a", both patterns eval("...") and eval("a") match eval(z).

def test(arg):
   if arg is not None:
      x = arg
      y = "1"
      z = "a"
      x = "v2"
      y = "2"
      z = "a"
   # ruleid: eval_arbitrary
   # OK
   # OK

Generic pattern matching

See generic pattern matching.