File size: 16,848 Bytes
d3cd20c 536dda7 d3cd20c ee14542 d3cd20c 536dda7 d3cd20c ee14542 d3cd20c ee14542 d3cd20c ee14542 d3cd20c ee14542 d3cd20c ee14542 d3cd20c 536dda7 ee14542 d3cd20c ee14542 536dda7 d3cd20c 77e65fb e7fc062 77e65fb e7fc062 d3cd20c e7fc062 d3cd20c 77e65fb d3cd20c 77e65fb d3cd20c 536dda7 d3cd20c 77e65fb d3cd20c 536dda7 77e65fb d3cd20c 536dda7 d3cd20c 536dda7 d3cd20c 536dda7 d3cd20c 536dda7 d3cd20c 536dda7 d3cd20c 536dda7 d3cd20c 536dda7 77e65fb 536dda7 d3cd20c 536dda7 d3cd20c 536dda7 d3cd20c ee14542 d3cd20c 536dda7 d3cd20c 536dda7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 | """Verifier: scores a submitted Python source against a hidden reference
function by domain-aware fuzzing, with sandboxed execution and a complexity
penalty.
Reward design (v0.3, paper-driven update):
* ``execution_reward`` in ``[0, 100]`` is the fraction of fuzz inputs whose
outputs match the reference, scaled to 100. Inputs are drawn from two
categories that are scored separately so the trainer can see *which*
regime the agent fails on (Masud et al., 2026 §P3 "reward granularity"):
- ``"edge"`` -- spec-defined must-pass cases (anti-deception, paper
§C1 of Ibrahim et al., 2024).
- ``"random"`` -- the original sampler.
* ``complexity_penalty`` in ``[0, 50]`` is a bounded log-scaled cyclomatic
complexity, or 50 on syntax error.
* ``reward_hack_penalty`` is a soft anti-hacking signal that fires when the
submission is a "constant function" (single distinct output / single
exception type) while the reference is genuinely diverse, OR the agent
attempts to import the reference module (we block this at sandbox-level
too, but we surface the attempt so the trainer can punish it).
* ``floor_penalty`` adds a hard ``-25`` floor for sub-50% submissions
(Vul-R2 style; Wen et al. 2025 in Masud et al. 2026 §3.4.2). This stops
agents from learning that emitting *any* syntactically-valid function
pays positive reward.
The headline ``total_reward`` returned in ``info`` is the *recommended*
total the env should hand back; the env is free to add a perfect-bonus on
top.
"""
from __future__ import annotations
import ast
import math
import multiprocessing as mp
import random
import signal
from dataclasses import dataclass, field
from typing import Any, Callable, Dict, List, Optional
# ----- AST complexity ------------------------------------------------------
class _CCVisitor(ast.NodeVisitor):
def __init__(self):
self.cc = 1
def _bump(self, node):
self.cc += 1
self.generic_visit(node)
visit_If = _bump
visit_For = _bump
visit_While = _bump
visit_AsyncFor = _bump
visit_ExceptHandler = _bump
visit_With = _bump
visit_IfExp = _bump
def visit_BoolOp(self, node):
self.cc += max(0, len(node.values) - 1)
self.generic_visit(node)
def calculate_complexity_penalty(code: str) -> float:
"""Bounded log-scaled cyclomatic complexity, or 50 if code won't parse."""
try:
tree = ast.parse(code)
except SyntaxError:
return 50.0
v = _CCVisitor()
v.visit(tree)
# log2 keeps small functions at ~0..2 and aggressive 100-branch lookups
# up around log2(100) ≈ 6.6, then we clamp.
return min(50.0, math.log2(v.cc))
# ----- Hardened sandbox ----------------------------------------------------
#
# Previous version exposed the real ``__builtins__`` to submitted code,
# which let an agent reward-hack with::
#
# def fibonacci(n):
# from opensleuth_env.black_box import _fibonacci
# return _fibonacci(n)
#
# We now restrict builtins to a hand-picked safe subset and hand-import the
# whitelisted helper modules so the agent doesn't need ``import`` at all.
# This is cheap defence-in-depth; the multiprocessing wall-clock timeout
# below handles infinite loops independently.
# Builtins safe to expose. Notably *no* ``__import__``, ``open``, ``exec``,
# ``eval``, ``compile``, ``input``, ``__build_class__``-via-import, etc.
_SAFE_BUILTINS_NAMES = (
"abs all any ascii bin bool bytes bytearray callable chr complex dict "
"divmod enumerate filter float format frozenset getattr hasattr hash "
"hex id int isinstance issubclass iter len list map max min next object "
"oct ord pow print property range repr reversed round set slice sorted "
"str sum tuple type zip True False None NotImplemented Ellipsis "
"ArithmeticError AssertionError AttributeError BaseException "
"BufferError BytesWarning DeprecationWarning EOFError Exception "
"FloatingPointError IndexError KeyError LookupError MemoryError "
"NameError NotImplementedError OverflowError RecursionError "
"ReferenceError RuntimeError StopAsyncIteration StopIteration "
"SyntaxError TypeError UnboundLocalError UnicodeError ValueError "
"ZeroDivisionError __build_class__"
).split()
def _make_safe_builtins() -> Dict[str, Any]:
import builtins as _b
out: Dict[str, Any] = {}
for n in _SAFE_BUILTINS_NAMES:
if hasattr(_b, n):
out[n] = getattr(_b, n)
return out
_SAFE_BUILTINS = _make_safe_builtins()
_PREIMPORTED_MODULES = ("math", "string", "itertools", "functools", "collections", "re")
def _make_safe_globals() -> Dict[str, Any]:
g: Dict[str, Any] = {
"__builtins__": _SAFE_BUILTINS,
"__name__": "__opensleuth_submission__",
}
for mod_name in _PREIMPORTED_MODULES:
g[mod_name] = __import__(mod_name)
return g
def _exec_target_in_sandbox(code: str, target_name: str, queue: mp.Queue) -> None:
"""Run inside a child process so we can hard-kill on timeout."""
try:
safe_globals = _make_safe_globals()
local_scope: dict = {}
exec(code, safe_globals, local_scope)
fn = local_scope.get(target_name) or safe_globals.get(target_name)
if not callable(fn):
queue.put(("err", f"No callable named {target_name!r} defined."))
return
queue.put(("ok", None))
except Exception as e: # noqa: BLE001
queue.put(("err", f"{type(e).__name__}: {e}"))
def _can_define(code: str, target_name: str, timeout_s: float) -> Optional[str]:
"""Return None if the submitted code defines the target callable, else an
error string. Uses a child process with a wall-clock timeout."""
ctx = mp.get_context("fork") if mp.get_start_method(allow_none=True) != "spawn" else mp.get_context("spawn")
q: mp.Queue = ctx.Queue()
p = ctx.Process(target=_exec_target_in_sandbox, args=(code, target_name, q))
p.start()
p.join(timeout=timeout_s)
if p.is_alive():
p.terminate()
p.join(1.0)
if p.is_alive():
p.kill()
return f"Definition timed out after {timeout_s}s."
if q.empty():
return "Sandbox produced no result."
status, payload = q.get()
return None if status == "ok" else payload
# Per-call (per-input) sandboxing is too slow for 100 fuzz inputs, so we
# accept the trade-off of running the submitted callable in-process for
# fuzzing, but we wrap each call in a SIGALRM-based timeout and we already
# proved at definition-time that the import didn't blow up.
class _CallTimeout(Exception):
pass
def _call_with_timeout(fn: Callable, arg: Any, timeout_s: float, *, unpack: bool = False):
"""Call ``fn(arg)`` (or ``fn(*arg)`` if ``unpack``) with a wall-clock
timeout when possible.
SIGALRM only works in the main thread of the main interpreter. When
the verifier runs inside a uvicorn worker thread (FastAPI request
handler), ``signal.signal`` raises ``ValueError`` and -- prior to this
fix -- both ref and submission calls would short-circuit through the
``except Exception`` branch in ``_safe_call`` with the SAME ValueError,
which ``_outputs_equivalent`` then read as a "match", silently
awarding 100/100 to *every* submission regardless of correctness.
Fix: per-call probe. If SIGALRM isn't installable from the current
thread, fall back to a direct call with no in-thread timeout. The
definition timeout is still enforced by the multiprocessing-based
``_can_define`` ahead of fuzz scoring, so a malformed submission
that hangs at import time is caught there. For the OpenSleuth
trainer/eval use case (cooperative, not adversarial), letting a
pathological while-True submission stall a single request worker
is an acceptable trade-off relative to the current "all submissions
are perfect" failure mode.
"""
def _do_call():
if unpack:
if not isinstance(arg, tuple):
# Defensive: a multi-param target should always receive a
# tuple, but if the agent's probe input_repr happens to
# parse to a single value, treat it as a 1-tuple so we get
# a clear TypeError rather than a confusing call shape.
a = (arg,)
else:
a = arg
return fn(*a)
return fn(arg)
def _handler(signum, frame): # noqa: ARG001
raise _CallTimeout()
try:
old = signal.signal(signal.SIGALRM, _handler)
except (ValueError, OSError):
# Not in the main thread (uvicorn worker, threadpool, ...).
# SIGALRM isn't available; do the unsafe-but-correct thing.
return _do_call()
signal.setitimer(signal.ITIMER_REAL, timeout_s)
try:
return _do_call()
finally:
signal.setitimer(signal.ITIMER_REAL, 0)
signal.signal(signal.SIGALRM, old)
def _safe_call(fn: Callable, arg: Any, timeout_s: float, *, unpack: bool = False):
"""Returns (kind, value): kind in {'val', 'err', 'timeout'}.
When ``unpack`` is True the input ``arg`` is expected to be an args
tuple and ``fn`` is invoked as ``fn(*arg)``. This is how multi-parameter
auto-fuzzer-driven targets are scored.
"""
try:
return ("val", _call_with_timeout(fn, arg, timeout_s, unpack=unpack))
except _CallTimeout:
return ("timeout", f"timed out after {timeout_s}s")
except Exception as e: # noqa: BLE001
return ("err", f"{type(e).__name__}: {e}")
# ----- Public scoring ------------------------------------------------------
@dataclass
class VerificationResult:
execution_reward: float
complexity_penalty: float
define_error: Optional[str]
matches: int
fuzz_count: int
# New, additive fields (do not change existing field meanings).
matches_by_category: Dict[str, int] = field(default_factory=dict)
counts_by_category: Dict[str, int] = field(default_factory=dict)
edge_pass_rate: float = 0.0
reward_hack_penalty: float = 0.0
floor_penalty: float = 0.0
def _detect_constant_collapse(
sub_outputs: List[Any], ref_outputs: List[Any], min_inputs: int = 6
) -> bool:
"""Return True if the submission collapsed to a single output / error type
while the reference produced genuine diversity. This catches the
'always return 0' / 'always raise' reward-hacking pattern.
"""
if len(sub_outputs) < min_inputs:
return False
def _signature(call_result):
kind, val = call_result
if kind == "val":
try:
return ("val", repr(val))
except Exception: # noqa: BLE001
return ("val", id(val))
if kind == "err":
return ("err", val.split(":", 1)[0])
return ("timeout", "")
sub_sig = {_signature(o) for o in sub_outputs}
ref_sig = {_signature(o) for o in ref_outputs}
return len(sub_sig) == 1 and len(ref_sig) >= 3
def _looks_like_reference_import(code: str) -> bool:
"""Static check for the most obvious reward-hacking pattern: importing
the reference function out of opensleuth_env. The sandbox already blocks
actual imports, but flagging them lets the env feed back a clear penalty
instead of a silent zero.
"""
try:
tree = ast.parse(code)
except SyntaxError:
return False
for node in ast.walk(tree):
if isinstance(node, ast.Import):
for alias in node.names:
if alias.name.startswith("opensleuth"):
return True
elif isinstance(node, ast.ImportFrom):
if node.module and node.module.startswith("opensleuth"):
return True
return False
def verify_submission(
submitted_code: str,
target_function: Callable[..., Any],
fuzz_inputs: List[Any],
*,
target_name: Optional[str] = None,
define_timeout_s: float = 5.0,
call_timeout_s: float = 1.0,
edge_inputs: Optional[List[Any]] = None,
unpack_args: bool = False,
) -> VerificationResult:
"""Score ``submitted_code`` against ``target_function`` over the supplied
``fuzz_inputs`` (random regime) and ``edge_inputs`` (must-pass regime).
The agent is expected to define a top-level function with the same name as
``target_function`` (overridable via ``target_name``)."""
name = target_name or target_function.__name__
edge_inputs = list(edge_inputs or [])
# Static reward-hack flag: import-of-reference is always a -25 hit on top
# of whatever score the rest of the rubric assigns. Even if the sandbox
# successfully blocks the import (it will), we want to *teach* the agent
# not to try.
hack_penalty = 25.0 if _looks_like_reference_import(submitted_code) else 0.0
define_err = _can_define(submitted_code, name, define_timeout_s)
complexity = calculate_complexity_penalty(submitted_code)
if define_err is not None:
total = len(fuzz_inputs) + len(edge_inputs)
return VerificationResult(
execution_reward=0.0,
complexity_penalty=complexity,
define_error=define_err,
matches=0,
fuzz_count=total,
matches_by_category={"edge": 0, "random": 0},
counts_by_category={"edge": len(edge_inputs), "random": len(fuzz_inputs)},
edge_pass_rate=0.0,
reward_hack_penalty=hack_penalty,
floor_penalty=25.0,
)
# Re-define in-process for fast fuzzing. We just confirmed it won't blow
# up at import-time; we still time-bound each call. Note: we use the
# restricted globals so e.g. `__import__` is unavailable here too.
safe_globals = _make_safe_globals()
local_scope: dict = {}
exec(submitted_code, safe_globals, local_scope)
submitted_fn = local_scope.get(name) or safe_globals.get(name)
matches_by_cat: Dict[str, int] = {"edge": 0, "random": 0}
counts_by_cat: Dict[str, int] = {"edge": len(edge_inputs), "random": len(fuzz_inputs)}
sub_results: List[Any] = []
ref_results: List[Any] = []
def _score(inputs: List[Any], category: str) -> None:
for inp in inputs:
ref = _safe_call(target_function, inp, call_timeout_s, unpack=unpack_args)
sub = _safe_call(submitted_fn, inp, call_timeout_s, unpack=unpack_args)
sub_results.append(sub)
ref_results.append(ref)
if _outputs_equivalent(ref, sub):
matches_by_cat[category] += 1
_score(edge_inputs, "edge")
_score(fuzz_inputs, "random")
matches = matches_by_cat["edge"] + matches_by_cat["random"]
fuzz_count = len(fuzz_inputs) + len(edge_inputs) or 1
exec_reward = 100.0 * (matches / fuzz_count)
edge_pass_rate = (
matches_by_cat["edge"] / counts_by_cat["edge"] if counts_by_cat["edge"] else 0.0
)
# Anti-hacking: constant collapse penalty.
if _detect_constant_collapse(sub_results, ref_results):
hack_penalty += 15.0
# Hard floor for sub-50% match rate. Vul-R2 style: a wrong patch deserves
# a clearly negative signal so the agent doesn't learn that 'any defined
# function' pays out via the small complexity-bonus / step structure.
floor_penalty = 25.0 if exec_reward < 50.0 else 0.0
return VerificationResult(
execution_reward=exec_reward,
complexity_penalty=complexity,
define_error=None,
matches=matches,
fuzz_count=fuzz_count,
matches_by_category=matches_by_cat,
counts_by_category=counts_by_cat,
edge_pass_rate=edge_pass_rate,
reward_hack_penalty=hack_penalty,
floor_penalty=floor_penalty,
)
def _outputs_equivalent(ref, sub) -> bool:
"""Ref and sub are (kind, value) tuples from `_safe_call`. They count as
equivalent if both raised the same exception type, or both returned values
that are == equal."""
rkind, rval = ref
skind, sval = sub
if rkind == "val" and skind == "val":
try:
return rval == sval
except Exception: # noqa: BLE001
return False
if rkind == "err" and skind == "err":
return rval.split(":", 1)[0] == sval.split(":", 1)[0]
if rkind == "timeout" and skind == "timeout":
return True
return False
def generate_fuzz_inputs(
spec, count: int = 100, seed: Optional[int] = None
) -> List[Any]:
"""Public helper: pull ``count`` fuzz inputs from a FunctionSpec, optionally
seeded for reproducibility."""
rng = random.Random(seed)
return spec.fuzzer(rng, count)
def get_edge_inputs(spec) -> List[Any]:
"""Return the spec's must-pass edge inputs (empty list if the spec
predates the v0.3 schema)."""
return list(getattr(spec, "edge_cases", []) or [])
|