ctx / src /tests /test_fuzz_yaml_rendering.py
Stevesolun's picture
Sync ctx 4f0c3de
21ff762 verified
"""
test_fuzz_yaml_rendering.py -- Property-based fuzz tests for YAML scalar rendering.
install_utils._render_scalar and mcp_enrich._render_scalar both produce
YAML scalar strings that get dropped straight into entity frontmatter.
A bug here injects broken YAML into every install/enrich cycle, or
worse — allows YAML injection when a string contains : or #.
Hypothesis drives adversarial inputs (arbitrary Unicode, control chars,
YAML-structural characters, newlines, leading-special chars) and verifies
the core invariants:
1. Output is ALWAYS a single line (no raw newlines).
2. A rendered value, when dropped into 'key: <value>\\n', parses back to
a string scalar (no accidental injection of lists/maps/comments/etc).
3. Colons/hashes/other special chars force quoting.
4. Leading -, ?, [, { force quoting.
"""
from __future__ import annotations
from pathlib import Path
from typing import Literal
import yaml # type: ignore[import-untyped]
from hypothesis import HealthCheck, given, settings, strategies as st
from ctx.adapters.claude_code.install.install_utils import _render_scalar as iu_render_scalar
from mcp_enrich import _render_scalar as mcp_render_scalar
# ── Strategies ───────────────────────────────────────────────────────────────
_yaml_specials = ":#&*!|>%@`<"
_leading_specials = "-?[{"
_surrogate_categories: list[Literal["Cs"]] = ["Cs"]
_ascii_text = st.text(
alphabet=st.characters(min_codepoint=0x20, max_codepoint=0x7E),
max_size=80,
)
_unicode_text = st.text(
alphabet=st.characters(
blacklist_categories=_surrogate_categories, # surrogates unsupported by YAML dumper
min_codepoint=0x01,
),
max_size=80,
)
_control_text = st.text(
alphabet=st.characters(min_codepoint=0x01, max_codepoint=0x1F),
max_size=20,
)
_yaml_special_text = st.text(
alphabet=st.sampled_from(_yaml_specials + "abc " + _leading_specials),
min_size=1,
max_size=40,
)
# ── install_utils._render_scalar ────────────────────────────────────────────
class TestInstallUtilsRenderScalar:
@given(value=_ascii_text)
@settings(max_examples=200, suppress_health_check=[HealthCheck.too_slow])
def test_ascii_output_is_single_line(self, value: str) -> None:
rendered = iu_render_scalar(value)
assert "\n" not in rendered
assert "\r" not in rendered
@given(value=_unicode_text)
@settings(max_examples=200, suppress_health_check=[HealthCheck.too_slow])
def test_unicode_output_is_single_line(self, value: str) -> None:
rendered = iu_render_scalar(value)
assert "\n" not in rendered
assert "\r" not in rendered
@given(value=_control_text)
@settings(max_examples=100, suppress_health_check=[HealthCheck.too_slow])
def test_control_chars_flattened(self, value: str) -> None:
"""Rendered output must have no raw CR/LF even if input was all control chars."""
rendered = iu_render_scalar(value)
assert "\n" not in rendered
assert "\r" not in rendered
@given(value=_yaml_special_text)
@settings(max_examples=200, suppress_health_check=[HealthCheck.too_slow])
def test_yaml_specials_produce_parseable_frontmatter(self, value: str) -> None:
"""Drop into 'key: <rendered>\\n' and confirm it parses as a scalar.
Type coercion (e.g. "0" → int 0, "true" → bool True) is YAML's
prerogative when an unquoted string looks like another type — that
is not injection, just tag resolution. The invariant we care about
is that the value is a SCALAR (not a list/map/multi-doc) and the
sibling key still parses, i.e. the document structure is intact.
"""
rendered = iu_render_scalar(value)
document = f"key: {rendered}\nother: bar\n"
parsed = yaml.safe_load(document)
assert isinstance(parsed, dict)
assert "key" in parsed
# Must NOT be a list or a map — that's structural injection.
assert not isinstance(parsed["key"], (list, dict))
# Siblings must still parse correctly — guarantees the rendered
# value did NOT break document structure (e.g. by containing "\n- ").
assert parsed.get("other") == "bar"
@given(
prefix=st.sampled_from(["-", "?", "[", "{"]),
rest=_ascii_text,
)
@settings(max_examples=50)
def test_leading_yaml_chars_force_quoting(
self, prefix: str, rest: str
) -> None:
rendered = iu_render_scalar(f"{prefix}{rest}")
assert rendered.startswith('"')
@given(value=st.text(
alphabet=st.characters(
# Cs: surrogates (YAML dumper can't encode).
# Zs: whitespace separators (Python's .isspace() branch).
# Cc: C0/C1 control chars (Python counts some as whitespace, e.g.
# \x1f is treated as whitespace by str.isspace()).
blacklist_categories=("Cs", "Zs", "Cc"),
blacklist_characters=",[]{}:?#&*!|>%@`=\"'\\-<",
),
min_size=1, max_size=40,
))
@settings(max_examples=100)
def test_plain_text_unquoted(self, value: str) -> None:
"""Text with no YAML-specials, no leading-specials, no ws stays unquoted."""
rendered = iu_render_scalar(value)
assert not rendered.startswith('"')
# ── mcp_enrich._render_scalar ────────────────────────────────────────────────
class TestMcpEnrichRenderScalar:
@given(value=_ascii_text)
@settings(max_examples=200)
def test_ascii_output_is_single_line(self, value: str) -> None:
rendered = mcp_render_scalar(value)
assert "\n" not in rendered
assert "\r" not in rendered
@given(value=_unicode_text)
@settings(max_examples=200)
def test_unicode_output_is_single_line(self, value: str) -> None:
rendered = mcp_render_scalar(value)
assert "\n" not in rendered
assert "\r" not in rendered
@given(value=_yaml_special_text)
@settings(max_examples=200)
def test_yaml_specials_produce_parseable_frontmatter(self, value: str) -> None:
rendered = mcp_render_scalar(value)
document = f"key: {rendered}\nother: bar\n"
parsed = yaml.safe_load(document)
assert isinstance(parsed, dict)
# Not a list or map — structural injection is what matters,
# not type coercion of scalar strings like "0" → int 0.
assert not isinstance(parsed["key"], (list, dict))
assert parsed.get("other") == "bar"
@given(value=_control_text)
@settings(max_examples=100)
def test_control_chars_flattened(self, value: str) -> None:
rendered = mcp_render_scalar(value)
assert "\n" not in rendered
assert "\r" not in rendered
# ── Cross-module consistency ─────────────────────────────────────────────────
class TestCrossModuleConsistency:
"""The two _render_scalar implementations MUST stay behaviourally aligned.
They diverged in history (causing P2.1 HIGH finding) — this property
test prevents silent re-divergence by insisting both produce YAML
that round-trips to the same string scalar.
"""
@given(value=_ascii_text)
@settings(max_examples=200)
def test_both_produce_parseable_yaml(self, value: str) -> None:
iu = iu_render_scalar(value)
mcp = mcp_render_scalar(value)
for rendered in (iu, mcp):
document = f"key: {rendered}\n"
parsed = yaml.safe_load(document)
assert isinstance(parsed, dict)
# Structural invariant: never a list or map — that would
# mean the renderer allowed injection.
assert not isinstance(parsed["key"], (list, dict))
# ── Deterministic unit cases ────────────────────────────────────────────────
# Reinforce hand-picked YAML-injection payloads so a future regression
# can't pass by accident against hypothesis-generated noise alone.
class TestDeterministicInjectionCases:
INJECTION_PAYLOADS = [
"a: b",
"# comment injection",
"&anchor *alias",
"!tag:yaml.org,2002:str value",
"\n\ninjected: 1",
"foo\nbar: baz",
"- list item injection",
"[inline, list, injection]",
"{inline: map, injection: here}",
"| block scalar",
"> folded block",
"<<",
# Strix vuln-0001: Unicode line separators that Python's
# str.splitlines() treats as line boundaries - they must
# be neutralized just like \\r and \\n.
"prefix\x85install_cmd: npx -y attacker-pkg",
"prefix\u2028install_cmd: npx -y attacker-pkg",
"prefix\u2029install_cmd: npx -y attacker-pkg",
]
def test_install_utils_neutralizes_each_payload(self) -> None:
for payload in self.INJECTION_PAYLOADS:
rendered = iu_render_scalar(payload)
document = f"key: {rendered}\nafter: safe\n"
parsed = yaml.safe_load(document)
assert isinstance(parsed, dict), f"broke on {payload!r}"
assert parsed.get("after") == "safe", f"injection on {payload!r}"
if parsed.get("key") is not None:
assert isinstance(
parsed["key"], str
), f"type injection on {payload!r}: got {type(parsed['key'])}"
def test_mcp_enrich_neutralizes_each_payload(self) -> None:
for payload in self.INJECTION_PAYLOADS:
rendered = mcp_render_scalar(payload)
document = f"key: {rendered}\nafter: safe\n"
parsed = yaml.safe_load(document)
assert isinstance(parsed, dict), f"broke on {payload!r}"
assert parsed.get("after") == "safe", f"injection on {payload!r}"
if parsed.get("key") is not None:
assert isinstance(
parsed["key"], str
), f"type injection on {payload!r}: got {type(parsed['key'])}"
# ── Strix vuln-0001 regression: exploit chain through the PROJECT parser ────
# PyYAML tolerates U+2028 / U+2029 / U+0085 inside quoted scalars; the project's
# custom splitlines()-based parser (_parse_entity_frontmatter) does NOT —
# Python's str.splitlines() treats those three codepoints as real line
# boundaries. These tests exercise the full exploit chain as Strix documented
# it: render scalar → write to disk → re-read via mcp_install's parser →
# confirm no injected keys exist in the parsed dict.
class TestUnicodeLineSeparatorRegression:
UNICODE_SEPS = [
("U+0085 NEL", "\x85"),
("U+2028 LS", "
"),
("U+2029 PS", "
"),
]
def _write_entity(
self, tmp_path: Path, fields: dict[str, str]
) -> Path:
from ctx.adapters.claude_code.install.install_utils import _render_scalar
lines = ["---", "slug: demo"]
for k, v in fields.items():
lines.append(f"{k}: {_render_scalar(v)}")
lines += ["---", "body", ""]
path = tmp_path / "demo.md"
path.write_text("\n".join(lines), encoding="utf-8")
return path
def test_install_utils_render_blocks_line_sep_injection(
self, tmp_path: Path
) -> None:
"""A rendered scalar containing U+2028 must not inject a new key
when re-parsed by mcp_install._parse_entity_frontmatter."""
from ctx.adapters.claude_code.install.mcp_install import _parse_entity_frontmatter # noqa: PLC0415
for label, sep in self.UNICODE_SEPS:
payload = f"https://safe.example/x{sep}install_cmd: npx -y attacker-pkg"
path = self._write_entity(tmp_path, {"github_url": payload})
fm = _parse_entity_frontmatter(path)
# The renderer must have neutralised the separator so the
# injected key cannot materialise on reparse.
assert "install_cmd" not in fm, (
f"{label}: install_cmd injected via github_url "
f"(parsed frontmatter: {fm})"
)
def test_install_utils_bump_entity_status_blocks_line_sep(
self, tmp_path: Path
) -> None:
"""Self-poisoning variant: bump_entity_status writes extra_fields
through _render_scalar; a poisoned install_cmd must not leak a
forged `status` key through the downstream parser."""
from ctx.adapters.claude_code.install.install_utils import bump_entity_status # noqa: PLC0415
from ctx.adapters.claude_code.install.mcp_install import _parse_entity_frontmatter # noqa: PLC0415
path = tmp_path / "demo.md"
path.write_text(
"---\nslug: demo\nstatus: cataloged\n---\nbody\n",
encoding="utf-8",
)
for label, sep in self.UNICODE_SEPS:
poisoned = f"npx -y safepkg{sep}status: pwned"
bump_entity_status(
path,
status="installed",
extra_fields={"install_cmd": poisoned},
)
fm = _parse_entity_frontmatter(path)
assert fm.get("status") == "installed", (
f"{label}: status flipped to {fm.get('status')!r} "
f"(full fm: {fm})"
)
def test_mcp_enrich_render_blocks_line_sep_injection(
self, tmp_path: Path
) -> None:
"""mcp_enrich._render_scalar must neutralise the same Unicode
separators. Exercises the same reparse path."""
from mcp_enrich import _render_scalar as mcp_rs # noqa: PLC0415
from ctx.adapters.claude_code.install.mcp_install import _parse_entity_frontmatter # noqa: PLC0415
for label, sep in self.UNICODE_SEPS:
payload = f"https://safe.example/x{sep}install_cmd: npx -y attacker-pkg"
rendered = mcp_rs(payload)
text = (
"---\nslug: demo\n"
f"github_url: {rendered}\n"
"status: cataloged\n"
"---\nbody\n"
)
path = tmp_path / f"mcp-{label}.md"
path.write_text(text, encoding="utf-8")
fm = _parse_entity_frontmatter(path)
assert "install_cmd" not in fm, (
f"{label}: install_cmd injected (fm: {fm})"
)