Security context
UnratedPYSEC-2022-43015 CVE-2022-45907Published Nov 26, 2022

In PyTorch before trunk/89695, torch.jit.annotations.parse_type_line can cause arbitrary code execution because eval is used unsafely.

Research this vulnerability

Research is free — Hunters explains how the bug works, the root-cause code pattern, how the fix addresses it, and how to test whether a target is affected, in chat. Investigate & write exploit is a paid run — the engine reads the advisory and fix commits, then builds and validates a working proof-of-concept exploit with reproduction steps.

Affected versions

0 → fixed in 1.13.1

Details

In PyTorch before trunk/89695, torch.jit.annotations.parse_type_line can cause arbitrary code execution because eval is used unsafely.

The fix

[JIT][Security] Do not blindly eval input string (#89189)

Nikita Shulga· Nov 17, 2022, 10:05 PM+213767f6aa49f
test/test_jit.py+8 0
@@ -3951,6 +3951,14 @@ def invalid4(a):
return a + 2
torch.jit.script(invalid4)
+ def test_calls_in_type_annotations(self):
+ with self.assertRaisesRegex(RuntimeError, "Type annotation should not contain calls"):
+ def spooky(a):
+ # type: print("Hello") -> Tensor # noqa: F723
+ return a + 2
+ print(torch.__file__)
+ torch.jit.annotations.get_signature(spooky, None, 1, True)
+
def test_is_optional(self):
ann = Union[List[int], List[float]]
torch._jit_internal.is_optional(ann)
torch/csrc/jit/frontend/script_type_parser.cpp+1 1
@@ -316,7 +316,7 @@ std::vector<IValue> ScriptTypeParser::evaluateDefaults(
// We then run constant prop on this graph and check the results are
// constant. This approach avoids having to have separate handling of
// default arguments from standard expressions by piecing together existing
- // machinery for graph generation, constant propgation, and constant
+ // machinery for graph generation, constant propagation, and constant
// extraction.
auto tuple_type = Subscript::create(
r,
torch/jit/annotations.py+12 2
@@ -1,4 +1,5 @@
import ast
+import dis
import enum
import inspect
import re
@@ -144,6 +145,15 @@ def check_fn(fn, loc):
raise torch.jit.frontend.FrontendError(loc, "Expected a single top-level function")
+def _eval_no_call(stmt, glob, loc):
+ """Evaluate statement as long as it does not contain any method/function calls"""
+ bytecode = compile(stmt, "", mode="eval")
+ for insn in dis.get_instructions(bytecode):
+ if "CALL" in insn.opname:
+ raise RuntimeError(f"Type annotation should not contain calls, but '{stmt}' does")
+ return eval(bytecode, glob, loc) # type: ignore[arg-type] # noqa: P204
+
+
def parse_type_line(type_line, rcb, loc):
"""Parses a type annotation specified as a comment.
@@ -154,7 +164,7 @@ def parse_type_line(type_line, rcb, loc):
arg_ann_str, ret_ann_str = split_type_line(type_line)
try:
- arg_ann = eval(arg_ann_str, {}, EvalEnv(rcb)) # type: ignore[arg-type] # noqa: P204
+ arg_ann = _eval_no_call(arg_ann_str, {}, EvalEnv(rcb))
except (NameError, SyntaxError) as e:
raise RuntimeError("Failed to parse the argument list of a type annotation") from e
@@ -162,7 +172,7 @@ def parse_type_line(type_line, rcb, loc):
arg_ann = (arg_ann,)
try:
- ret_ann = eval(ret_ann_str, {}, EvalEnv(rcb)) # type: ignore[arg-type] # noqa: P204
+ ret_ann = _eval_no_call(ret_ann_str, {}, EvalEnv(rcb))
except (NameError, SyntaxError) as e:
raise RuntimeError("Failed to parse the return type of a type annotation") from e

References