This should not work:-

python3.7 -c  'print(eval("().__class__.__base__.__subclasses__()[-1].__init__.__globals__",{"__builtins__": {}},{"__builtins__": {}}))'

and should be properly fixed.

Would you care to explain why this should not work and how this behavior is in violation of the language specification?

It is perfectly valid expression. From a security perspective it may be an undesired feature. However Python does neither claim nor promise that eval is secure, see articel https://lwn.net/Articles/574215/ for more information on a failed attempt to sandbox Python. There is also ast.literal_eval() function, which provides limit evaluation of simple expressions.

I suggest you to read https://python-security.readthedocs.io/security.html about the Python security model. In short, as soon as you let users to execute arbitrary Python code, they get a full access to the machine.

If you want to restrict access, you must run Python inside a restricted container (or any sandbox).

The specification specifically allows for the restriction of access to globals via the second argument to eval.

While Christian and Victor make interesting, albeit suicidal, comments and references to other efforts, the fact remains that this is a violation of the standard, and is an exploitable security issue.

It's worth noting that the 1980's are long over now - people take security seriously these days, even when it's inconvenient.

The fix seems ridiculously trivial for what it's worth; introduce a flag that honors the intent of the second argument.

Your assumption is incorrect. The eval() does not promise that default builtins cannot be access through other means. The behavior has been discussed several times and at great length over the past decade.

> The specification specifically allows for the restriction of access to globals via the second argument to eval.

The Python language reference doesn't provide any warranty like that.
https://docs.python.org/dev/library/functions.html#eval

I close the issue as "not as bug".

All previous attempts to "sandbox" Python code in Python have failed. The correct way is to run Python in a sandbox. Not the opposite.

> https://lwn.net/Articles/574215/ 

This one was my attempt for example ;-)