fix(aura): reject history-free agents by default (#2652)

* fix(aura): reject history-free agents by default

* fix(aura): keep invalid responses fail-closed
This commit is contained in:
haelyra 2026-08-02 13:42:16 -04:00 committed by GitHub
parent e4e4163101
commit f782bd616e
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 96 additions and 32 deletions

View file

@ -18,7 +18,7 @@ from aura import before_settle, AuraUntrusted
def settle(counterparty_did: str, amount: float) -> None: def settle(counterparty_did: str, amount: float) -> None:
try: try:
before_settle(counterparty_did) # rejects high_risk + unknown before_settle(counterparty_did) # rejects high_risk + new + unknown
except AuraUntrusted as e: except AuraUntrusted as e:
log.warning("blocked: %s", e) log.warning("blocked: %s", e)
return # your policy decides what to do return # your policy decides what to do
@ -41,9 +41,8 @@ if v.dimensions and v.dimensions.get("financial_integrity", 1) < 0.4:
require_manual_review() # placeholder for your own policy require_manual_review() # placeholder for your own policy
``` ```
> `v.ok` reflects the *verdict class* (True for `trusted`/`caution`), not the > `v.ok` reflects the *verdict class* (True for `trusted`/`caution`). Use the
> outcome of `require_trust()` — the gate's default `allow` also lets `new` > gate's return/raise for the policy decision and `v.ok` for display.
> through. Use the gate's return/raise for the decision, `v.ok` for display.
## Verdicts ## Verdicts
@ -58,8 +57,8 @@ if v.dimensions and v.dimensions.get("financial_integrity", 1) < 0.4:
## Policy knobs ## Policy knobs
```python ```python
# Reject brand-new agents too (strict): # Explicitly allow brand-new agents during a controlled onboarding flow:
before_settle(did, allow=("trusted", "caution")) before_settle(did, allow=("trusted", "caution", "new"))
# Treat an *unreachable* AURA as a pass (fail-open). Off by default — # Treat an *unreachable* AURA as a pass (fail-open). Off by default —
# absence of evidence is not evidence of trust. # absence of evidence is not evidence of trust.
@ -78,11 +77,15 @@ before_settle(did, base_url="https://my-aura-mirror.example", timeout=5)
- **default (`fail_open=False`)**`unknown` is rejected → an unreachable AURA - **default (`fail_open=False`)**`unknown` is rejected → an unreachable AURA
blocks the action. *Fail-closed.* blocks the action. *Fail-closed.*
- **`fail_open=True`** — `unknown` from an unreachable endpoint is allowed - **`new` verdict** — rejected by default because the agent has no interaction
through, so AURA can never take your flow down. *Fail-open.* history. Onboarding flows can explicitly add `new` to `allow`.
- **`fail_open=True`** — `unknown` from a transport failure is allowed through.
HTTP errors, malformed JSON, and invalid response shapes remain blocked
because the endpoint was reached but did not return a trustworthy verdict.
This keeps the trust signal **purely additive**: if you remove the adapter or Removing the adapter leaves your existing allow/deny logic untouched. While
AURA is down, your existing allow/deny logic runs exactly as before. the gate is enabled, an AURA outage blocks the protected action by default;
callers must explicitly choose `fail_open=True` to preserve availability.
## Tests ## Tests

View file

@ -10,10 +10,9 @@ Design boundary (intentional):
- read-only: the only network call is GET /check?did=... - read-only: the only network call is GET /check?did=...
- no auth: /check is a public endpoint; no API key, no secret - no auth: /check is a public endpoint; no API key, no secret
- no coupling: pure stdlib (urllib). No third-party imports, no SDK. - no coupling: pure stdlib (urllib). No third-party imports, no SDK.
- fail-closed: on network failure the verdict is `unknown`, and the - fail-closed: by default, the gate rejects agents without interaction
default gate (before_settle) rejects `unknown` so an history (`new`) and agents it cannot verify (`unknown`).
unreachable AURA never silently waves a counterparty Flip `fail_open=True` to excuse transport failures only.
through. Flip `fail_open=True` to invert that.
Public API: Public API:
aura_verdict(did) -> AuraVerdict (never raises on network) aura_verdict(did) -> AuraVerdict (never raises on network)
@ -43,9 +42,10 @@ __all__ = [
DEFAULT_BASE_URL = "https://agent.auraopenprotocol.org" DEFAULT_BASE_URL = "https://agent.auraopenprotocol.org"
DEFAULT_TIMEOUT = 8 # seconds DEFAULT_TIMEOUT = 8 # seconds
# Verdicts safe to proceed with by default. Rejects `high_risk` (poor track # Verdicts safe to proceed with by default. `new` remains available as an
# record) and `unknown` (no verifiable history / endpoint unreachable). # explicit opt-in for onboarding flows, but history-free agents should not
DEFAULT_ALLOW = ("trusted", "caution", "new") # satisfy a reputation gate automatically.
DEFAULT_ALLOW = ("trusted", "caution")
# All verdict classes the /check endpoint can return. # All verdict classes the /check endpoint can return.
VERDICTS = ("trusted", "caution", "high_risk", "new", "unknown") VERDICTS = ("trusted", "caution", "high_risk", "new", "unknown")
@ -82,10 +82,10 @@ class AuraVerdict:
score: Optional[float] = None score: Optional[float] = None
has_history: bool = False has_history: bool = False
dimensions: Optional[dict[str, float]] = None dimensions: Optional[dict[str, float]] = None
# False only when AURA could not be reached (network/parse failure) and the # False only when AURA could not be reached because of a transport failure.
# verdict is a synthetic `unknown`. A reachable AURA that genuinely returns # HTTP errors, malformed JSON, invalid shapes, and genuine `unknown`
# `unknown` has reachable=True. before_settle's fail_open keys on this, not # verdicts remain reachable=True. before_settle's fail_open keys on this,
# on the verdict alone, so it can't wave through unverified counterparties. # not on the verdict alone, so it cannot wave through invalid responses.
reachable: bool = True reachable: bool = True
raw: dict[str, Any] = field(default_factory=dict, repr=False) raw: dict[str, Any] = field(default_factory=dict, repr=False)
@ -121,9 +121,14 @@ class AuraVerdict:
@classmethod @classmethod
def unreachable(cls, did: str, reason: str) -> "AuraVerdict": def unreachable(cls, did: str, reason: str) -> "AuraVerdict":
"""A synthetic `unknown` verdict for network/parse failures.""" """A synthetic `unknown` verdict for transport failures."""
return cls(did=did, verdict="unknown", reason=reason, reachable=False) return cls(did=did, verdict="unknown", reason=reason, reachable=False)
@classmethod
def invalid_response(cls, did: str, reason: str) -> "AuraVerdict":
"""A reachable endpoint response that could not be trusted."""
return cls(did=did, verdict="unknown", reason=reason, reachable=True)
# Indirection point so tests can inject canned responses without a network. # Indirection point so tests can inject canned responses without a network.
# Signature: (url: str, timeout: float) -> dict (raises on transport error) # Signature: (url: str, timeout: float) -> dict (raises on transport error)
@ -156,13 +161,15 @@ def aura_verdict(
url = f"{base_url.rstrip('/')}/check?" + urllib.parse.urlencode({"did": did}) url = f"{base_url.rstrip('/')}/check?" + urllib.parse.urlencode({"did": did})
try: try:
body = _fetch(url, timeout) body = _fetch(url, timeout)
except urllib.error.HTTPError as e:
return AuraVerdict.invalid_response(did, f"AURA returned HTTP {e.code}: {e.reason}")
except (urllib.error.URLError, TimeoutError, OSError) as e: except (urllib.error.URLError, TimeoutError, OSError) as e:
return AuraVerdict.unreachable(did, f"AURA unreachable: {e}") return AuraVerdict.unreachable(did, f"AURA unreachable: {e}")
except (json.JSONDecodeError, ValueError) as e: except (json.JSONDecodeError, ValueError) as e:
return AuraVerdict.unreachable(did, f"AURA returned non-JSON: {e}") return AuraVerdict.invalid_response(did, f"AURA returned non-JSON: {e}")
if not isinstance(body, dict): if not isinstance(body, dict):
return AuraVerdict.unreachable(did, "AURA returned an unexpected shape") return AuraVerdict.invalid_response(did, "AURA returned an unexpected shape")
return AuraVerdict.from_payload(did, body) return AuraVerdict.from_payload(did, body)
@ -180,13 +187,13 @@ def before_settle(
raises AuraUntrusted on fail. raises AuraUntrusted on fail.
try: try:
before_settle(counterparty_did) # rejects high_risk + unknown before_settle(counterparty_did) # rejects high_risk + new + unknown
settle_payment(counterparty_did, amount) settle_payment(counterparty_did, amount)
except AuraUntrusted as e: except AuraUntrusted as e:
abort(str(e)) abort(str(e))
Tighten to reject brand-new agents too: Explicitly allow brand-new agents in an onboarding flow:
before_settle(did, allow=("trusted", "caution")) before_settle(did, allow=("trusted", "caution", "new"))
fail_open=True makes an *unreachable* AURA pass through (transport failure fail_open=True makes an *unreachable* AURA pass through (transport failure
only a reachable AURA that returns `unknown` is still rejected). Off by only a reachable AURA that returns `unknown` is still rejected). Off by

View file

@ -13,6 +13,8 @@ Coverage:
from __future__ import annotations from __future__ import annotations
import json
from typing import Any
import urllib.error import urllib.error
import pytest import pytest
@ -70,9 +72,14 @@ def test_gate_allows_trusted():
assert v.verdict == "trusted" assert v.verdict == "trusted"
def test_gate_allows_caution_and_new_by_default(): def test_gate_allows_caution_by_default() -> None:
assert before_settle("did:aura:caution-bot", _fetch=FETCH).verdict == "caution" assert before_settle("did:aura:caution-bot", _fetch=FETCH).verdict == "caution"
assert before_settle("did:aura:fresh-bot", _fetch=FETCH).verdict == "new"
def test_gate_rejects_new_by_default() -> None:
with pytest.raises(AuraUntrusted) as exc_info:
before_settle("did:aura:fresh-bot", _fetch=FETCH)
assert exc_info.value.verdict.verdict == "new"
def test_gate_rejects_high_risk(): def test_gate_rejects_high_risk():
@ -86,9 +93,13 @@ def test_gate_rejects_unknown_by_default():
before_settle("did:aura:ghost-bot", _fetch=FETCH) before_settle("did:aura:ghost-bot", _fetch=FETCH)
def test_strict_allow_rejects_new(): def test_opt_in_allow_can_include_new() -> None:
with pytest.raises(AuraUntrusted): v = before_settle(
before_settle("did:aura:fresh-bot", allow=("trusted", "caution"), _fetch=FETCH) "did:aura:fresh-bot",
allow=("trusted", "caution", "new"),
_fetch=FETCH,
)
assert v.verdict == "new"
# ── network-failure path ────────────────────────────────────────────────────── # ── network-failure path ──────────────────────────────────────────────────────
@ -120,6 +131,49 @@ def test_fail_open_does_not_pass_reachable_unknown():
before_settle("did:aura:ghost-bot", fail_open=True, _fetch=FETCH) before_settle("did:aura:ghost-bot", fail_open=True, _fetch=FETCH)
def test_fail_open_does_not_pass_malformed_response() -> None:
fetch = raising_fetch(json.JSONDecodeError("expecting value", "<html>", 0))
with pytest.raises(AuraUntrusted) as exc_info:
before_settle(
"did:aura:trusted-bot",
fail_open=True,
_fetch=fetch,
)
assert exc_info.value.verdict.reachable is True
def test_fail_open_does_not_pass_invalid_response_shape() -> None:
def invalid_shape_fetch(_url: str, _timeout: float) -> Any:
return []
with pytest.raises(AuraUntrusted) as exc_info:
before_settle(
"did:aura:trusted-bot",
fail_open=True,
_fetch=invalid_shape_fetch,
)
assert exc_info.value.verdict.reachable is True
def test_fail_open_does_not_pass_http_error_response() -> None:
fetch = raising_fetch(
urllib.error.HTTPError(
"https://agent.auraopenprotocol.org/check",
503,
"service unavailable",
None,
None,
)
)
with pytest.raises(AuraUntrusted) as exc_info:
before_settle(
"did:aura:trusted-bot",
fail_open=True,
_fetch=fetch,
)
assert exc_info.value.verdict.reachable is True
def test_reachable_verdict_marked_reachable(): def test_reachable_verdict_marked_reachable():
v = aura_verdict("did:aura:ghost-bot", _fetch=FETCH) v = aura_verdict("did:aura:ghost-bot", _fetch=FETCH)
assert v.reachable is True assert v.reachable is True