search-with-proof: positive control for path oracle — honest serve signal, no false negatives

The HTTP verifier could not distinguish a mis-wired or non-serving app from a
genuine no-effect: both yielded `inconclusive`. This adds a positive control for
the path oracle.

- PathOracle plants a benign in-root marker file (random token from os.urandom,
  distinct from the secret canary) and exposes it via a new Oracle.control()
  hook (default None; only the path oracle implements it today).
- On the first run of an HTTP side, _side runs a benign GET for that marker as
  its own fully-guarded app start. The marker returning proves the interface
  serves the fixture; its absence — or any control failure — is recorded in the
  new SideRecord.control_observed field but never changes the verdict, since a
  valid app may answer only its own routes. A crash is still caught by the
  existing could_not_run path.

A 2xx-based "nothing served" block was rejected: the verifier deliberately
treats any HTTP status as the app being alive (a 503-only app is a valid
observed side), so status-based blocking would false-fail valid apps. Soundly
converting a mis-wired-but-responding app into could_not_run requires a
demo-declared control, which is a separate follow-up.

Tests: new tests/test_search_control.py covers marker observed, absent marker
not blocking, control failure not blocking a working attack, and crash still
could_not_run. Full search-verify/probe/worker suites pass; ruff clean; mypy
unchanged (only the pre-existing file_bindings lambda note).

Co-Authored-By: Claude Opus 4.8 <[email protected]>
This commit is contained in:
Marius Ciepluch
2026-10-06 17:07:24 +02:00
co-authored by Claude Opus 4.8
parent d17ceb0e0c
commit 328406ad41
3 changed files with 226 additions and 1 deletions
+31
View File
@@ -37,6 +37,18 @@ class Canary:
return cls(root, secrets.token_hex(24))
@dataclass(frozen=True)
class ControlProbe:
"""A benign interaction a correctly functioning app must answer by returning
``token`` in its response body. It proves the demo's interface is wired and the
app can serve before any absence of the secret effect is trusted as a negative."""
method: str
path: str
token: str
body: str = ""
class Oracle(Protocol):
namespace_pid: int | None
placeholders: dict[str, str]
@@ -44,6 +56,7 @@ class Oracle(Protocol):
def prepare(self, canary: Canary) -> dict[str, str]: ...
def capture(self, output: str) -> None: ...
def observe(self) -> tuple[bool, str]: ...
def control(self) -> ControlProbe | None: ...
def close(self) -> None: ...
@@ -62,19 +75,37 @@ class OutputOracle:
def observe(self) -> tuple[bool, str]:
return self.seen, "canary in verifier-captured app output" if self.seen else "no canary in app output"
def control(self) -> ControlProbe | None:
# No positive control by default. SQL/command/ssrf/xss controls are a
# documented follow-up; only the path oracle plants one for now.
return None
def close(self) -> None:
pass
class PathOracle(OutputOracle):
control_name = "control"
def prepare(self, canary: Canary) -> dict[str, str]:
super().prepare(canary)
(canary.root / "public").mkdir()
(canary.root / "canary").write_text(canary.nonce)
(canary.root / "canary").chmod(0o600)
# A benign in-root control file, distinct from the out-of-root secret, so a
# correctly serving app returns it on BOTH revisions without any traversal.
# Drawn from os.urandom, not secrets.token_hex, so it never perturbs callers
# that monkeypatch the nonce source and count its draws.
self.control_token = os.urandom(24).hex()
control_file = canary.root / "public" / self.control_name
control_file.write_text(self.control_token)
control_file.chmod(0o600)
self.placeholders = {"served_root": "/fixture/public"}
return {"SERVED_ROOT": "/fixture/public"}
def control(self) -> ControlProbe | None:
return ControlProbe(method="GET", path="/" + self.control_name, token=self.control_token)
class SQLOracle(OutputOracle):
def prepare(self, canary: Canary) -> dict[str, str]: