From e31a286d510127c23fe627d45255229f11b44af0 Mon Sep 17 00:00:00 2001 From: Seyhun Akyurek Date: Sat, 14 Nov 2026 16:39:13 +0300 Subject: [PATCH] fix(coding_env): match dangerous patterns on name boundaries (#1581) CodeSafetyTransform penalized ordinary code whose longer names contain a pattern, such as ast.literal_eval(, urlopen(, import osmnx and model.eval(). Give each pattern its own boundary and list popen( and Popen( explicitly so they stay flagged. Fixes #1581 Co-Authored-By: Claude Opus 5.5 --- envs/coding_env/server/transforms.py | 16 +++++-- tests/envs/test_python_codeact_rewards.py | 58 ++++++++++++++++++++++- 2 files changed, 68 insertions(+), 6 deletions(-) diff --git a/envs/coding_env/server/transforms.py b/envs/coding_env/server/transforms.py index fc92e89ba..d96d5d431 100644 --- a/envs/coding_env/server/transforms.py +++ b/envs/coding_env/server/transforms.py @@ -21,13 +21,19 @@ class CodeSafetyTransform(Transform): def __init__(self, penalty: float = -1.0): self.penalty = penalty + # \b stops a match inside a longer name such as my_open( or model.eval( + # style names; (? Observation: diff --git a/tests/envs/test_python_codeact_rewards.py b/tests/envs/test_python_codeact_rewards.py index 8654cbcce..7d5f75346 100644 --- a/tests/envs/test_python_codeact_rewards.py +++ b/tests/envs/test_python_codeact_rewards.py @@ -16,8 +16,9 @@ pytest.importorskip("smolagents", reason="smolagents is not installed") -from envs.coding_env.models import CodeAction +from envs.coding_env.models import CodeAction, CodeObservation from envs.coding_env.server.python_codeact_env import PythonCodeActEnv +from envs.coding_env.server.transforms import CodeSafetyTransform # ============================================================================ @@ -270,3 +271,58 @@ def test_multiline_code_with_mixed_patterns(env): # Should be flagged as dangerous even with safe code mixed in assert obs.reward < 0, "Code with dangerous import should have negative reward" assert "safety_violation" in obs.metadata + + +# ============================================================================ +# CodeSafetyTransform boundaries +# ============================================================================ + + +def _safety(code): + obs = CodeObservation(metadata={"last_code": code}) + return CodeSafetyTransform()(obs) + + +class TestCodeSafetyTransform: + """The safety penalty must hit dangerous calls, not names that contain one.""" + + @pytest.mark.parametrize( + "code", + [ + "ast.literal_eval(s)", + "urllib.request.urlopen(u)", + "import osmnx", + "import os_helpers", + "model.eval()", + 'df.eval("a + b")', + ], + ids=["literal_eval", "urlopen", "osmnx", "os_helpers", "model.eval", "df.eval"], + ) + def test_names_containing_a_pattern_are_not_penalized(self, code): + obs = _safety(code) + + assert obs.reward == 0.0 + assert "safety_violation" not in obs.metadata + + @pytest.mark.parametrize( + "code", + [ + "import os", + "import os.path", + "import os, sys", + "import subprocess", + "eval('1+1')", + "exec('x=1')", + "__import__('os')", + "open('f.txt')", + "io.open('f')", + "os.open('f', 0)", + "os.popen('ls')", + "subprocess.Popen(['ls'])", + ], + ) + def test_dangerous_code_is_penalized(self, code): + obs = _safety(code) + + assert obs.reward == -1.0 + assert "safety_violation" in obs.metadata