Evalidate is simple python module for safe and very fast eval()'uating user-supplied (possible malicious) python expressions.
Version 2.0 is backward incompatible with older versions. safeeval()
and evalidate()
methods are removed, and EvalMode class is introduced.
See upgrade example in ticket or use older (any before 2.0.0, e.g. v1.1.0) if you have old code and do not want to upgrade. But upgrading is easy, so please consider this option.
Originally it's developed for filtering complex data structures e.g.
Find cheap smartphones available for sale:
category=="smartphones" and price<300 and stock>0
But also, it can be used for other expressions, e.g. arithmetical, like
a+b-100
Evalidate is fastest among all (known to me) secure eval python modules.
pip3 install evalidate
Built-in python features such as compile() or eval() are quite powerful to run any kind of user-supplied code, but could be insecure if used code is malicious like os.system("rm -rf /")
. Evalidate works on whitelist principle, allowing code only if it consist only of safe operations (based on authors views about what is safe and what is not, your mileage may vary - but you can supply your list of safe operations)
from evalidate import Expr, EvalException
src = 'a + 40 > b'
# src = "__import__('os').system('clear')"
try:
print(Expr(src).eval({'a':10, 'b':42}))
except EvalException as e:
print(e)
Gives output: True
In case of dangerous code (uncomment second src line to test):
output will be: ERR: Operation type Call is not allowed
Evalidate throws exceptions CompilationException
, ValidationException
, ExecutionException
. All of them
inherit from base exception class EvalException
.
Evalidate is very flexible, depending on security model, same code can either pass validation or raise exception.
EvalModel is security model class for eval - lists of allowed AST nodes, function calls, attributes and dict of imported functions. There is built-in model base_eval_model
with basic operations allowed (which are safe from authors point of view).
You can create custom empty model (and extend it later):
my_model = evalidate.EvalModel()
(nothing is allowed by default, even 1+2
will not be considered safe)
or you may start from base_eval_mode
and extend it:
from evalidate import Expr, base_eval_model
my_model = base_eval_model.clone()
my_model.nodes.append('Mult')
Expr('2*2', model=my_model).eval()
To enable int()
function, need to allow 'Call'
node and add this function to list of allowed function:
my_model.nodes.append('Call')
my_model.allowed_functions.append('int')
Expr('int(36.6)', model=my_model).eval()
Or, to call attributes:
m = base_eval_model.clone()
m.nodes.extend(['Call', 'Attribute'])
m.attributes.append('startswith')
src = '"abcdef".startswith("abc")'
r = evalidate.Expr(src, model=m).eval()
But even with this settings, exploiting it with expression like __builtins__["eval"](1)
will fail (good!).
def one():
return 1
m = base_eval_model.clone()
m.nodes.append('Call')
m.imported_functions["one"] = one
Expr('one()', model=m).eval()
Evalidate is very fast, but it's still takes CPU cycles... If you want to achieve maximal possible speed, you can use python native eval with this kind of code:
from evalidate import Expr
d = dict(a=1, b=2)
expr = Expr('a+b')
eval(expr.code, None, d) # <-- native python eval, will run at eval() speed
This is as secure as expr.eval(), because expr.code
is already validated to be secure.
Difference is very little: execution of expr.code
can throw any exception, while expr.eval()
can throw only ExecutionException. Also, if you want to export your functions to eval, you should do this manually.
evalidate uses ast.parse() to get AST node to validate it.
Warning
It is possible to crash the Python interpreter with a sufficiently large/complex string due to stack depth limitations in Python’s AST compiler.
In my test, works well with 200 nested int(): int(int(.... int(1)...))
but not with 201. Source code is 1000+ characters. But even if evalidate will get such code, it will just raise CompilationException
.
Evalidate is very flexible and it's possible to shoot yourself in foot if you will try hard. test_security()
checks your configuration (nodes, funcs, attrs) against given list of possible attack code or against built-in list of attacks. test_security()
returns True if everything is OK (all attacks raised ValidationException) or False if something passed.
This code will never print (I hope).
from evalidate.security import test_security
test_security() or print("default rules are vulnerable!")
But this will fail because nodes/funcs leads to successful validation for attack (suppose you do not want anyone to call int()
)
from evalidate.security import test_security
attacks = ['int(1)']
test_security(attacks, addnodes=['Call'], funcs=['int'], verbose=True)
It will print:
Testing attack code:
int(1)
Problem! Attack passed validation without exception!
Code:
int(1)
This is code of examples/products.py
. Expression is validated and compiled once and executed (as byte-code, very fast) many times, so filtering is both fast and secure.
#!/usr/bin/env python3
import requests
from evalidate import Expr, ValidationException, CompilationException, ExecutionException
import json
import sys
data = requests.get('https://dummyjson.com/products?limit=100').json()
try:
src = sys.argv[1]
except IndexError:
src = 'True'
try:
expr = Expr(src)
except (ValidationException, CompilationException) as e:
print(e)
sys.exit(1)
c=0
for p in data['products']:
# print(p)
try:
r = expr.eval(p)
if r:
print(json.dumps(p, indent=2))
c+=1
except ExecutionException as e:
print("Runtime exception:", e)
print("# {} products matches".format(c))
# print all 100 products
./products.py
# Only cheap products, 8 matches
./products.py 'price<20'
# smartphones (5)
./products.py 'category=="smartphones"'
# good smartphones
./products.py 'category=="smartphones" and rating>4.5'
# cheap smartphones
./products.py 'category=="smartphones" and price<300'
While asteval can compute much more complex code (define functions, use python math libraries) it has drawbacks:
- asteval is much slower (evalidate can be used at speed of eval() python bytecode)
- user can provide source code which runs very long time and consumes many resources
simpleeval
Very similar project, using AST approach too and optimized to re-evaluate pre-parsed expressions. But parsed expressions are stored as more high-level ast.Expr type and this approach is few times slower, while evalidate uses python native code
type and evaluation itself goes at speed of python eval()
evalidate is good to run same expression against different data.
We use benchmark/benchmark.py
in this repository.
We prepare list of 1 million of products (actually, we take just 100 products sample, but repeat it 10 000 times to get 1 million), and then filter it, finding only specific products on "untrusted user-supplied expression" (price < 20
in this case)
Products: 1000000 items
evalidate_raw_eval(): 0.266s
evalidate_eval(): 0.326s
test_simpleeval(): 1.824s
test_asteval(): 26.106s
As you see, evalidate is few times faster then simpleeval and both are much faster then asteval.
Maybe my test is not perfectly optimized (I'm not expert with simpleeval/asteval), if you can suggest better filtering sample code (which produces faster result), I will include it. (Benchmark code must assume expression as unknown in advance and untrusted)
- https://nedbatchelder.com/blog/201206/eval_really_is_dangerous.html
- https://netsec.expert/posts/breaking-python3-eval-protections/
- https://realpython.com/python-eval-function/
Note: realpython article shows example with nice short method of validation source (using code.co_names
),
but it's vulnerable, it passes "bomb" from Ned Batchelder article (bomb has empty co_names
tuple) and crash interpreter. Evalidate can block this code and similar bombs (unless you will intentionally configure evalidate to pass specific bomb code. Yes, with evalidate it is hard to shoot yourself in the foot, but it is possible if you will try hard).
Want more info? Check source code of module, it's very short and simple, easy to modify
Write me: yaroslaff at gmail.com