secure-torch

Model Trust Gateway — Unified, defense-in-depth Python library for safely loading AI model files.

What is secure-torch?

Loading third-party AI models is a supply chain risk. torch.load, ONNX Runtime, and safetensors have all been vectors for RCE, data exfiltration, and silent model tampering.

secure-torch is a Model Trust Enforcement Layer — not just a safe loader wrapper.

It adds what the ML ecosystem is missing:

How it complements safetensors

safetensors solves unsafe deserialization and memory safety. secure-torch adds the trust layer on top — signature verification, publisher policy, and provenance validation — for safetensors, PyTorch pickle, and ONNX alike.

Quick Start

pip install secure-torch

Drop-in usage

import secure_torch as torch          # drop-in replacement

model = torch.load("model.pt")        # safe by default

Require a valid signature

model = torch.load(
    "model.pt",
    require_signature=True,
    bundle_path="model.pt.sigstore",  # online Rekor verification
)

Offline verification (enterprise / air-gapped)

model = torch.load(
    "model.pt",
    require_signature=True,
    pubkey_path="trusted.pub",        # Ed25519 public key
    bundle_path="model.pt.sig",       # raw signature bytes
)

Trusted publishers allowlist

model = torch.load(
    "model.pt",
    trusted_publishers=["huggingface.co/meta", "openai.com"],
)

Audit first, block later (gradual adoption)

model, report = torch.load("model.pt", audit_only=True)

print(report.threat_level)       # ThreatLevel.LOW
print(report.score_breakdown)    # {'unsigned_model': 40}
print(report.warnings)           # ['No signature bundle found']

Sandbox isolation

model = torch.load("model.pt", sandbox=True)
# Model loaded in restricted subprocess (strict exec/network blocking via seccomp on Linux)

Other compatibility surfaces

torch.jit.load("model.pt")  # secure pipeline for local artifacts
torch.hub.load("pytorch/vision", "resnet50")      # remote convenience passthrough
torch.from_pretrained("bert-base-uncased")        # remote convenience passthrough
torch.save(model, "model.pt")

torch.hub.load and direct torch.from_pretrained calls do not currently enforce secure-torch security checks for remote fetches. Passing security arguments (require_signature, trusted_publishers, audit_only, max_threat_score, sandbox, sbom_*, bundle_path, pubkey_path) raises SecurityError. For enforced security controls, download artifacts first and call torch.load(local_path, ...), or use secure_torch.patch_huggingface(...) for Hugging Face download interception.

Hugging Face integration (automatic download scanning)

import secure_torch
from transformers import AutoModel

# Global monkey-patch: intercept hf_hub_download and validate model files.
secure_torch.patch_huggingface(require_signature=False, max_threat_score=20)

model = AutoModel.from_pretrained("gpt2")

# Optional cleanup when done.
secure_torch.unpatch_huggingface()

patch_huggingface(...) hooks huggingface_hub.file_download.hf_hub_download. For model-like files (.pt, .pth, .bin, .safetensors, .onnx), secure-torch runs format detection, validators, signature/publisher checks, and threat policy evaluation before returning the downloaded file path. Unsafe artifacts are blocked with UnsafeModelError.

Interactive CLI audit

# Styled report (rich-enabled when available)
secure-torch audit model.pt

# Machine-readable output
secure-torch audit model.pt --json

Pipeline

Every secure_load() call runs this fixed pipeline — steps cannot be skipped:

secure_load(file)
  1. format detect        ← .safetensors / .pt / .onnx / magic bytes
  2. signature verify     ← fail fast if require_signature=True
  3. threat score         ← explainable named dict
  4. policy enforce       ← trusted_publishers check
  5. sandbox load         ← subprocess (+ seccomp on Linux)
  6. return tensors

Threat Scoring

Scores are named and explainable — not a magic number:

model, report = torch.load("model.pt", audit_only=True)
print(report.score_breakdown)
# {
#   'unsigned_model': 40,
#   'custom_ops_detected': 30,
#   'unknown_publisher': 20,
# }
# total: 90 → ThreatLevel.CRITICAL

Default max_threat_score is 20 (MEDIUM). Adjust with max_threat_score=N.

SBOM Policy (Experimental)

Note: SPDX AI Profile is not yet widely adopted (early 2026). This feature is forward-looking — secure-torch is helping establish the standard.

from secure_torch.sbom.spdx_parser import parse_sbom
from secure_torch.sbom.opa_runner import OPAPolicyRunner

sbom = parse_sbom("model.spdx.json")
runner = OPAPolicyRunner("policy/production.rego")
denials = runner.evaluate(sbom, context={"environment": "production"})

if denials:
    raise RuntimeError(f"Policy violations: {denials}")

Example policy (policy/production.rego):

package secure_torch.policy

deny[msg] {
    input.sensitivePersonalInformation == "yes"
    msg := "Model contains sensitive personal information"
}

deny[msg] {
    ds := input.aiProfile.trainingDatasets[_]
    startswith(ds.license, "GPL")
    input.environment == "production"
    msg := sprintf("GPL dataset '%v' blocked in production", [ds.name])
}

Supported Formats

Signature Verification

Mode 1 — Online (Sigstore / Rekor)

# Sign your model
cosign sign-blob model.pt --bundle model.pt.sigstore

# Verify on load
torch.load("model.pt", require_signature=True, bundle_path="model.pt.sigstore")

Mode 2 — Offline (Ed25519 pubkey)

# Sign
openssl genpkey -algorithm ed25519 -out private.pem
openssl pkey -in private.pem -pubout -out public.pem
openssl pkeyutl -sign -inkey private.pem -in model.pt -out model.pt.sig

# Verify on load
torch.load("model.pt", require_signature=True, pubkey_path="public.pem", bundle_path="model.pt.sig")

CVE Coverage

Installation

pip install secure-torch

# With ONNX support
pip install secure-torch[onnx]

# With Sigstore online verification
pip install secure-torch[sigstore]

# With offline public-key verification
pip install secure-torch[crypto]

# Everything
pip install secure-torch[all]

Development

git clone https://github.com/Avanishk05/secure-torch
cd secure-torch
pip install -e .
pip install pytest pytest-cov mypy ruff bandit
pytest tests/

License

Apache 2.0 — see LICENSE.