
Proof-of-concept exploit for CVE-2026-47627, a path traversal vulnerability in NVIDIA Triton Inference Server leading to arbitrary file write via Zip-Slip. Includes detailed analysis, lab environment, and one-click exploit scripts.
Complete research writeup & PoC in this repo (image
nvcr.io/nvidia/tritonserver:26.05-py3/server v2.69.0). Repo is set up to always beexplicitso PoC works without restart via gRPC.
| Field | Nilai |
|---|---|
| CVE | CVE-2026-47627 |
| CNA | [email protected] |
| Publish | 2026-08-18 (NVD Awaiting Analysis) |
| Bulletin | NV 5865 → https://github.com/NVIDIA/product-security/tree/main/2026/5865 |
| CVSS 3.1 | 9.8 CRITICAL AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H |
| CWE | CWE-22 Path Traversal |
| Affected | 0.0-26.05 (server v2.69.0 / core r26.05) |
| Fixed | 26.06 (server v2.70.0 / core r26.06) |
| Impact | DoS (deskripsi CNA) — riset ini membuktikan file write outside /models via Zip-Slip, yang bisa diperluas ke DoS/resource exhaustion |
| Credit | Martin Brodeur |
Honest note: The CNA description is only
path traversal → DoSwithout details. This writeup reconstructs the bug location via diffr26.05..r26.06and direct testing in the26.05-py3container.
Image: nvcr.io/nvidia/tritonserver:26.05-py3 (TRITON_VERSION 2.69.0)
Compose is always explicit (already patched in this repo):
# docker-compose.yml:12
command: ["tritonserver", "--model-repository=/models", "--model-control-mode=explicit", "--log-verbose=0"]
run_triton.bat:10, run_triton.ps1:11, run_triton.sh:12 are also already explicit.
Triton Mode:
NONE (default upstream) → load once at start, POST /v2/repository/models/.../load → 503, must restart to load a new model.POLL → scan repo every second, auto-load, still blocks explicit load.EXPLICIT (this repo) → no scanning, but opens gRPC load_model / POST /load → PoC without restart via gRPC.Port:
| 8000 | HTTP REST | v2/health, v2/models, v2/repository |
| 8001 | gRPC | RepositoryModelLoad, ModelInfer |
| 8002 | Metrics | Prometheus |
Initial model:
model_repository/
├── echo_python/ (python backend, model.py)
│ └── 1/model.py
└── identity_onnx/ (onnxruntime, model.onnx ~1KB)
Quickstart:
docker compose up -d
docker compose logs -f
py scripts/client_test.py # health + infer identity_onnx
User input (model_name / tar entry) ──► join(parent, child) ──► canonicalize ──► cek "child di dalam parent?" ──► buka file
│ │ │ │ │
└──► "../tmp/pwn" /models + "/../tmp/pwn" = /models/../tmp/pwn → /tmp/pwn rfind vs find + '/' check FileExists / extract
join is used without canonicalize or with an incorrect rfind(parent,0)==0 check (partial match /models vs /models_evil), then ../ can escape./dev/zero, /proc/self/mem, or a tar containing ../../ that overwrites system files → hang / OOM / crash.Three research questions (from the initial report):
model_name in the gRPC RepositoryModelLoad, EXECUTION_ENV_PATH tar entry, file: override, TRITON_BATCH_STRATEGY_PATH.posixpath.join(repo, model_name) / temp_dir + "/" + file_name + IsChildPathEscapingParentPath.DoS via resource exhaustion / hang. In this repo it is proven to cause a file write to /tmp/poc_marker.Repo core (r26.05..r26.06):
src/filesystem/api.cc:407 IsChildPathEscapingParentPath — fix dcb315d + 72f3d9b:
// VULN (r26.05):
absolute_child.rfind(absolute_parent,0) != 0
// → bug partial-match: "/models" prefix dari "/models_evil/file" dianggap inside
// FIXED (r26.06):
canonical_child.find(canonical_parent,0)==0 &&
((child.size() > parent.size() && child[parent.size()]=='/') || child.size()==parent.size())
// → harus '/' setelah prefix, "/models_evil" sekarang correctly outside
Used in src/backend_model.cc:196 (TRITON_BATCH_STRATEGY_PATH) and src/model_repository_manager/model_repository_manager.cc:162 (file:).
src/model_repository_manager/model_repository_manager.cc:62 ValidateModelName — core#472/#481 (March 2026, already in 26.05 but tightened in 26.06):
if (trimmed==".." || trimmed.find('/')!=npos) return INVALID_ARG "must not contain path traversal"
This is what makes gRPC load_model('../tmp/pwn') → 400 INVALID_ARGUMENT in 26.05 (already blocked). But %2f (encoded /) slips through because it isn't a literal / → 500 literal poll.
python_be.cc:292 EXECUTION_ENV_PATH — e520f8c7 test zipslip_test.py:
The tar is extracted without checking for .. / absolute. The entry ../../poc_marker from /models/poc_exploit/malicious_env.tar.gz is extracted to /tmp/poc_marker (outside the model dir). Fix in 26.06: use ARCHIVE_EXTRACT_SECURE_NODOTDOT|NOABSOLUTEPATHS + check IsChildPathEscapingParentPath.
Repo server (r26.05..r26.06): only sagemaker_server.cc:1027 (RE2::FullMatch check) + http_server.cc:2440 (atoi→stoi), not the path for this CVE.
Conclusion: The model_name path via HTTP/gRPC is already blocked in 26.05 by ValidateModelName, but tar Zip-Slip is not — this is what the PoC in this repo exploits.
| Vector | Payload | 26.05 Result | Hit? |
|---|---|---|---|
HTTP raw ../ | POST /v2/repository/models/../tmp/pwn/load | 404 (routing normalize) | ❌ |
HTTP encoded ..%2f | POST /v2/repository/models/..%2ftmp%2fpwn/load | 500 literal, poll /models/..%2ftmp%2fpwn doesn't exist | ❌ (fake 500, not traversal) |
gRPC raw ../tmp/pwn | load_model('../tmp/pwn') | 400 INVALID_ARGUMENT | ❌ (already blocked) |
gRPC encoded ..%2f | load_model('..%2ftmp%2fpwn') | 500 literal poll | ❌ |
file: override ../models_evil | file:../models_evil/pwn | 400/500 depending on partial-match | ⚠️ (depends on IsChildPathEscapingParentPath bug, but already blocked by ValidateModelName for model_name) |
EXECUTION_ENV_PATH tar ../../poc_marker | malicious_env.tar.gz entry ../../poc_marker | 200 + file at /tmp/poc_marker | ✅ VULN |
The
500 vs 400analogy in the old writeup is wrong:500= file doesn't exist (no door),400= rejected by validation (door locked). Neither gets in. Only200+ a file at/tmpis proof.
Idea: The Python backend parameter EXECUTION_ENV_PATH points to $$TRITON_MODEL_DIRECTORY/malicious_env.tar.gz. When load_model('poc_exploit') is called, the backend pb_env.cc:292 extracts the tar without sanitization. The entry ../../poc_marker escapes from .../poc_exploit/ to /tmp/poc_marker.