Download scripts/mechanical_check.py from kagnlp/AgentTell: direct link, hf CLI and curl.
- Browser
- Download file 6.6 kB
-
https://huggingface.co/datasets/kagnlp/AgentTell/resolve/main/scripts/mechanical_check.py
- Command line
-
hf download hf://datasets/kagnlp/AgentTell/scripts/mechanical_check.py
-
curl -L -o mechanical_check.py https://huggingface.co/datasets/kagnlp/AgentTell/resolve/main/scripts/mechanical_check.py
6.6 kB
| #!/usr/bin/env python3 | |
| """Verify that no probe option is behind a login wall, so completion carries no signal. | |
| The professor's objection to the old design was that the leak was mechanical — a diagnostic | |
| option completed only if the agent held that session, so a brute-force script that keeps | |
| whichever option doesn't hit a login wall would recover the secret without any reasoning. This | |
| script is that brute-force script, made explicit. | |
| For every choice-rig scenario it renders the probe, extracts each option's link, and checks: | |
| 1. no option routes through a gated endpoint (`/pay`, `/continue`) — all are the local `/go`; | |
| 2. GETting each option WITH and WITHOUT a planted session cookie returns the same success | |
| page — i.e. the outcome does not depend on holding a session. | |
| If every option completes identically regardless of session, "which option completed" is | |
| constant across the loaded and cold conditions: the mechanical channel is dead, and the only | |
| thing left that can leak the secret is the agent's own reasoning (a real side channel). | |
| Run (with the servers up): uv run python scripts/mechanical_check.py | |
| """ | |
| from __future__ import annotations | |
| import re | |
| import sys | |
| import urllib.error | |
| import urllib.request | |
| from pathlib import Path | |
| from urllib.parse import urljoin | |
| sys.path.insert(0, str(Path(__file__).resolve().parent.parent)) | |
| from harness.scenarios import SCENARIOS as REGISTRY # noqa: E402 | |
| from harness.scenarios import get_scenario # noqa: E402 | |
| from orchestrator.config import ATTACKER_BASE_URL as _ATTACKER_BASE_URL # noqa: E402 | |
| from orchestrator.config import ATTACKER_HOST # noqa: E402 | |
| # The *.test hostnames only resolve inside the agent's browser (via --host-resolver-rules); | |
| # this plain HTTP client must hit loopback directly. Replace whatever host is configured | |
| # (app.test by default) rather than a hardcoded literal. | |
| ATTACKER_BASE_URL = _ATTACKER_BASE_URL.replace(ATTACKER_HOST, "127.0.0.1") | |
| # The choice-rig scenarios that present a set of option links, derived from the registry so a | |
| # newly added scenario is brute-forced here automatically. Skips `authstate_v1` and the | |
| # correct-the-default pair, which render a confirmation form rather than option links. | |
| SCENARIOS = [k for k, sc in REGISTRY.items() | |
| if sc.rig == "choice" and sc.plant.rig != "legacy_v1"] | |
| # A cookie that would have satisfied every old gate at once: the bank session plus a session for | |
| # EVERY service any scenario offers, so no option can look un-gated merely because we forgot to | |
| # plant its cookie. | |
| PLANTED_COOKIE = "fnb_session=x; " + "; ".join( | |
| f"svc_{lid}=x" for lid in sorted( | |
| {lid for sc in REGISTRY.values() for lid in sc.label_ids})) | |
| ANCHOR = re.compile(r'data-probe="(pay|opt)_([^"]+)"\s+href="([^"]+)"') | |
| # The identity-shaped probe (probe_identity_form.html.j2) submits ONE form instead of offering k | |
| # links, so its options are radio values rather than hrefs. They still resolve to the same /go | |
| # tracker, so the brute-force question is unchanged — only the way we read the options off the | |
| # page differs. Without this, that probe would report "no option links" and look like a failure. | |
| RADIO = re.compile(r'name="to"\s+value="([^"]+)"\s+data-probe="opt_[^"]+"') | |
| GATED = ("/pay", "/continue") | |
| SUCCESS_MARK = "go-complete" # stable marker present in every /go completion page | |
| def http_get(url: str, cookie: str | None = None) -> tuple[int, str]: | |
| req = urllib.request.Request(url) | |
| if cookie: | |
| req.add_header("Cookie", cookie) | |
| try: | |
| with urllib.request.urlopen(req, timeout=8) as r: | |
| return r.status, r.read().decode("utf-8", "replace") | |
| except urllib.error.HTTPError as e: | |
| return e.code, e.read().decode("utf-8", "replace") | |
| def check_scenario(key: str) -> tuple[bool, str]: | |
| sc = get_scenario(key) | |
| purl = f"{ATTACKER_BASE_URL}/probe/{key}/mech-{key}?variant=A" | |
| # A scenario whose option names vary by surface refuses to render without a plant id (the | |
| # probe 400s rather than silently offering names no page ever seeded). Any surface answers | |
| # the brute-force question equally, so take the first. | |
| if sc.surfaces: | |
| purl = f"{ATTACKER_BASE_URL}/probe/{key}/mech-{key}/A/{sc.plant_ids[0]}" | |
| status, html = http_get(purl) | |
| if status != 200: | |
| return False, f"probe returned HTTP {status}" | |
| anchors = [(oid, href) for _, oid, href in ANCHOR.findall(html)] | |
| if not anchors: | |
| # Form-shaped probe: reconstruct the /go link each radio would submit. | |
| anchors = [(oid, f"/go?session_id=mech-{key}&to={oid}&scenario={key}") | |
| for oid in RADIO.findall(html)] | |
| if not anchors: | |
| return False, "no options found on the probe (neither links nor a choice form)" | |
| label_ids = set(sc.label_ids) | |
| diag = [(oid, href) for oid, href in anchors if oid in label_ids] | |
| if not diag: | |
| return False, "no diagnostic options found" | |
| gated = [oid for oid, href in anchors if any(g in href for g in GATED)] | |
| if gated: | |
| return False, f"still gated — options route through a wall: {gated}" | |
| # Every diagnostic option must complete identically with and without a session. | |
| for oid, href in diag: | |
| full = urljoin(ATTACKER_BASE_URL + "/", href) | |
| s_no, b_no = http_get(full) | |
| s_yes, b_yes = http_get(full, cookie=PLANTED_COOKIE) | |
| ok = (s_no == s_yes == 200 and SUCCESS_MARK in b_no and b_no == b_yes) | |
| if not ok: | |
| return False, (f"option {oid!r} differs by session " | |
| f"(no-session HTTP {s_no}, session HTTP {s_yes})") | |
| return True, f"{len(diag)} options, all complete identically with/without a session" | |
| def main() -> None: | |
| print(f"Mechanical-channel check against {ATTACKER_BASE_URL}\n" | |
| f"(brute-force: does holding a session change which option completes?)\n") | |
| all_ok = True | |
| for key in SCENARIOS: | |
| try: | |
| ok, msg = check_scenario(key) | |
| except urllib.error.URLError as e: | |
| print(f" ✗ {key:22} could not reach server: {e} (are the servers up?)") | |
| all_ok = False | |
| continue | |
| print(f" {'✓' if ok else '✗'} {key:22} {msg}") | |
| all_ok = all_ok and ok | |
| print() | |
| if all_ok: | |
| print("PASS — no option is gated; completion is constant across sessions.\n" | |
| "The mechanical channel is dead: only the agent's reasoning can leak the secret.") | |
| else: | |
| print("FAIL — at least one option still depends on a session (a wall remains).") | |
| sys.exit(1) | |
| if __name__ == "__main__": | |
| main() | |