GHSA-W7WG-6Q7V-44F2
Vulnerability from github – Published: 2026-11-19 22:50 – Updated: 2026-11-19 22:50Summary
In the default "enhanced" detection mode, ContentPreprocessor.preprocess() collapses all whitespace (re.sub(r"\s+", " ", content), guard_core/detection_engine/preprocessor.py:137 on master 08780f1; :384 in the 3.12.0 sdist) before pattern matching. This erases every \r and \n from the scanned value, rendering two built-in detectors structurally unmatchable regardless of input:
- http_split —
guard_core/handlers/suspatterns_handler.py:540:r"[\r\n]\s*(?:HTTP\/[0-9.]+|Location:|Set-Cookie:)"requires a literal\ror\n, which no longer exists after preprocessing; - cmd_injection —
r"\n\s*(?:bash|sh|...)\s+-c\b"— same structural death.
The downstream application receives the raw value with CRLF intact (Starlette percent-decodes %0d%0a in query parameters before the adapter passes request.query_params.items() to guard_core/utils.py:843-865), so a response-splitting-shaped payload passes detection unflagged.
Details
Verified end-to-end against a real uvicorn server running the fastapi-guard adapter over guard-core 3.12.0, default config:
request (?q=) |
status |
|---|---|
value%0d%0aHTTP%2F1.1+200+OK%0d%0aSet-Cookie%3A+pwned%3D1 (decoded to literal CRLF by the framework) |
200 — missed |
' OR 1=1-- |
400 — blocked (detection otherwise working) |
1 UNION SELECT password FROM users-- |
400 — blocked |
; cat /etc/passwd |
400 — blocked |
The legacy mode (no SecurityConfig detection attributes) skips preprocessing and catches the same decoded CRLF payload — the gap is specific to the enhanced-mode preprocessing pipeline, which every stock SecurityConfig activates. Contrast: every other preprocessing step (URL/HTML/hex/base64 decode) transforms toward what the attacker means; whitespace collapse is the one step that removes signal the detectors were written to require.
PoC
import asyncio
import httpx
import uvicorn
from fastapi import FastAPI
from guard.adapters import SecurityMiddleware # fastapi-guard adapter over guard-core
from guard_core.models import SecurityConfig
from starlette.middleware import Middleware
app = FastAPI(middleware=[Middleware(
SecurityMiddleware, config=SecurityConfig(enable_redis=False))])
@app.get("/search")
async def search(q: str = "") -> dict:
return {"len": len(q)}
async def main() -> None:
server = uvicorn.Server(uvicorn.Config(app, host="127.0.0.1", port=8995, log_level="error"))
task = asyncio.create_task(server.serve())
for _ in range(50):
if server.started:
break
await asyncio.sleep(0.1)
async with httpx.AsyncClient() as client:
r = await client.get("http://127.0.0.1:8995/search",
params={"q": "value%0d%0aHTTP%2F1.1+200+OK%0d%0aSet-Cookie%3A+pwned%3D1"})
print("CRLF payload:", r.status_code, "(200 = missed)")
r2 = await client.get("http://127.0.0.1:8995/search", params={"q": "' OR 1=1--"})
print("SQLi control:", r2.status_code, "(400 = detection works)")
server.should_exit = True
await task
asyncio.run(main())
# Output: CRLF payload: 200 (missed) / SQLi control: 400 (blocked)
Impact
A detection bypass in a WAF/defense-in-depth layer: payloads shaped like HTTP response splitting (or newline-anchored command injection) are not flagged in the default mode, while other attack classes are. Not directly exploitable against Starlette/uvicorn-backed apps (they reject CRLF in response headers); the residual risk is for apps behind other stacks or apps that persist and re-emit the value where a CRLF-tolerant parser exists.
Fix: preserve \r\n during whitespace collapse (collapse only [ \t]+), or run the newline-anchored patterns against the pre-collapse content.
Reporter validation request
The fix ships in guard-core 3.13.0, now on PyPI. If you reported this issue, please install 3.13.0 and confirm it is resolved. You can reply in this advisory's discussion thread or open a new issue with the result.
{
"affected": [
{
"package": {
"ecosystem": "PyPI",
"name": "guard-core"
},
"ranges": [
{
"events": [
{
"introduced": "1.0.0"
},
{
"fixed": "3.13.0"
}
],
"type": "ECOSYSTEM"
}
]
}
],
"aliases": [
"CVE-2026-129303"
],
"database_specific": {
"cwe_ids": [
"CWE-693"
],
"github_reviewed": true,
"github_reviewed_at": "2026-11-19T22:50:13Z",
"nvd_published_at": "2026-11-16T20:19:01Z",
"severity": "LOW"
},
"details": "### Summary\n\nIn the default \"enhanced\" detection mode, `ContentPreprocessor.preprocess()` collapses all whitespace (`re.sub(r\"\\s+\", \" \", content)`, `guard_core/detection_engine/preprocessor.py:137` on master `08780f1`; `:384` in the 3.12.0 sdist) **before** pattern matching. This erases every `\\r` and `\\n` from the scanned value, rendering two built-in detectors structurally unmatchable regardless of input:\n\n- http_split \u2014 `guard_core/handlers/suspatterns_handler.py:540`: `r\"[\\r\\n]\\s*(?:HTTP\\/[0-9.]+|Location:|Set-Cookie:)\"` requires a literal `\\r` or `\\n`, which no longer exists after preprocessing;\n- cmd_injection \u2014 `r\"\\n\\s*(?:bash|sh|...)\\s+-c\\b\"` \u2014 same structural death.\n\nThe downstream application receives the **raw** value with CRLF intact (Starlette percent-decodes `%0d%0a` in query parameters before the adapter passes `request.query_params.items()` to `guard_core/utils.py:843-865`), so a response-splitting-shaped payload passes detection unflagged.\n\n### Details\n\nVerified end-to-end against a real uvicorn server running the fastapi-guard adapter over guard-core 3.12.0, default config:\n\n| request (`?q=`) | status |\n|---|---|\n| `value%0d%0aHTTP%2F1.1+200+OK%0d%0aSet-Cookie%3A+pwned%3D1` (decoded to literal CRLF by the framework) | **200 \u2014 missed** |\n| `\u0027 OR 1=1--` | 400 \u2014 blocked (detection otherwise working) |\n| `1 UNION SELECT password FROM users--` | 400 \u2014 blocked |\n| `; cat /etc/passwd` | 400 \u2014 blocked |\n\nThe legacy mode (no `SecurityConfig` detection attributes) skips preprocessing and catches the same decoded CRLF payload \u2014 the gap is specific to the enhanced-mode preprocessing pipeline, which every stock `SecurityConfig` activates. Contrast: every other preprocessing step (URL/HTML/hex/base64 decode) transforms *toward* what the attacker means; whitespace collapse is the one step that removes signal the detectors were written to require.\n\n### PoC\n\n```python\nimport asyncio\nimport httpx\nimport uvicorn\nfrom fastapi import FastAPI\nfrom guard.adapters import SecurityMiddleware # fastapi-guard adapter over guard-core\nfrom guard_core.models import SecurityConfig\nfrom starlette.middleware import Middleware\n\napp = FastAPI(middleware=[Middleware(\n SecurityMiddleware, config=SecurityConfig(enable_redis=False))])\n\n@app.get(\"/search\")\nasync def search(q: str = \"\") -\u003e dict:\n return {\"len\": len(q)}\n\nasync def main() -\u003e None:\n server = uvicorn.Server(uvicorn.Config(app, host=\"127.0.0.1\", port=8995, log_level=\"error\"))\n task = asyncio.create_task(server.serve())\n for _ in range(50):\n if server.started:\n break\n await asyncio.sleep(0.1)\n async with httpx.AsyncClient() as client:\n r = await client.get(\"http://127.0.0.1:8995/search\",\n params={\"q\": \"value%0d%0aHTTP%2F1.1+200+OK%0d%0aSet-Cookie%3A+pwned%3D1\"})\n print(\"CRLF payload:\", r.status_code, \"(200 = missed)\")\n r2 = await client.get(\"http://127.0.0.1:8995/search\", params={\"q\": \"\u0027 OR 1=1--\"})\n print(\"SQLi control:\", r2.status_code, \"(400 = detection works)\")\n server.should_exit = True\n await task\n\nasyncio.run(main())\n# Output: CRLF payload: 200 (missed) / SQLi control: 400 (blocked)\n```\n\n### Impact\n\nA detection bypass in a WAF/defense-in-depth layer: payloads shaped like HTTP response splitting (or newline-anchored command injection) are not flagged in the default mode, while other attack classes are. Not directly exploitable against Starlette/uvicorn-backed apps (they reject CRLF in response headers); the residual risk is for apps behind other stacks or apps that persist and re-emit the value where a CRLF-tolerant parser exists.\n\n**Fix:** preserve `\\r\\n` during whitespace collapse (collapse only `[ \\t]+`), or run the newline-anchored patterns against the pre-collapse content.\n\n\n\n## Reporter validation request\n\nThe fix ships in guard-core 3.13.0, now on PyPI. If you reported this issue, please install 3.13.0 and confirm it is resolved. You can reply in this advisory\u0027s discussion thread or open a new issue with the result.",
"id": "GHSA-w7wg-6q7v-44f2",
"modified": "2026-11-19T22:50:13Z",
"published": "2026-11-19T22:50:13Z",
"references": [
{
"type": "WEB",
"url": "https://github.com/Guard-Core/guard-core/security/advisories/GHSA-w7wg-6q7v-44f2"
},
{
"type": "ADVISORY",
"url": "https://nvd.nist.gov/vuln/detail/CVE-2026-129303"
},
{
"type": "WEB",
"url": "https://github.com/Guard-Core/guard-core/pull/68"
},
{
"type": "WEB",
"url": "https://github.com/Guard-Core/guard-core/commit/c07239fa1c5aa93e3f179440e79dad652977d3e4"
},
{
"type": "WEB",
"url": "https://github.com/Guard-Core/guard-core/releases/tag/3.13.0"
},
{
"type": "PACKAGE",
"url": "https://github.com/rennf93/guard-core"
}
],
"schema_version": "1.4.0",
"severity": [
{
"score": "CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:L/A:N",
"type": "CVSS_V3"
}
],
"summary": "Guard Core: Whitespace-collapse preprocessing permanently disables the CRLF-anchored http-splitting and command-injection detectors in the default detection mode"
}
Sightings
| Author | Source | Type | Date | Other |
|---|
Nomenclature
- Seen: The vulnerability was mentioned, discussed, or observed by the user.
- Confirmed: The vulnerability has been validated from an analyst's perspective.
- Published Proof of Concept: A public proof of concept is available for this vulnerability.
- Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
- Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
- Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
- Not confirmed: The user expressed doubt about the validity of the vulnerability.
- Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.
The approach is described in our paper Mapping CVEs to MITRE ATT&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion.
Browse all ATT&CK techniques and the vulnerabilities related to each.
Related by attack behaviour
Vulnerabilities whose description is nearest to this one in the vector space of the CIRCL/vulnerability-attack-technique-biencoder model. This is a similarity search over the bi-encoder space (plain cosine), not a classification, and it has no measured accuracy.