AcademyResourcesCompanyResearchBook a demo ↗
Platform / Test & Secure / Probexa
Model Risk Scanning
Probexa

Scan the model before you trust the model.

Probexa is a static security scanner for ML models — detecting unsafe serialization, remote-code-execution risks, malicious loaders and supply-chain weaknesses, all without ever executing the model.

Static analysis onlyNo model executionCI/CD readySARIF outputPolicy verdicts
Probexa · static scan✓ COMPLETE
fraud-model.zip · finance profile · no execution
Text Indicator
Pickle Gadget
TensorFlow
Keras Custom Layer
ONNX Analyzer
Remote Loader
Requirements
Archive
Trojan Heuristics
60%
Safe format rate
2
Remote loaders
3
Pickle gadgets
0.18
Malicious density
VerdictFAILmodel blocked from deployment
pass warn fail
/ The new blind spot
You scan your code. You scan your containers.
Are you scanning your models?
Internal data scienceOpen-source reposHugging Face & hubsVendors & partnersResearch teamsContractors
9
Specialized scanners in parallel
15+
Model-hygiene risk metrics
0
Model executions — static only
SARIF
2.1.0 output for code scanning
/ How it scans

Fast, safe, policy-driven.

Three principles make Probexa safe to run on untrusted models and easy to operationalize.

01

Static analysis only

Probexa scans model artifacts without executing them — safe for sensitive environments where unknown models must never be loaded into runtime.

02

Specialized ML coverage

Not a generic file scanner. Purpose-built scanners for ML formats, serialization, configs, dependencies, remote loaders and archives.

03

Policy-based decisions

Technical findings become clear Pass / Warn / Fail verdicts via configurable policy profiles for healthcare, finance and baseline use.

/ How Probexa works

One scanner. The whole model supply chain.

Point Probexa at any model — open-source, vendor or internal. It statically inspects every artifact across six detection families, then returns a policy verdict, ranked findings and CI-ready outputs.

Open Source ModelsHugging Face · PyTorch · TensorFlow· · ·Vendor ModelsAWS · Azure · Google Cloud· · ·Internal ModelsRegistry · Datasets · Artifacts· · ·ProbexaML Model ScannerModel Supply Chain SecurityUnsafeSerializationPickleGadgetsRemoteLoadersDependencyRisksArchiveTraversalTrojanHeuristics1Policy VerdictPassWarnFail2Metrics & Findings124FINDINGSCritical18High32Medium46Low28findings by scanner family3OutputsJSONSARIFCI/CDHealthcare · Finance · GlobalBuilt for regulated industriesand trusted worldwide.
/ How it works

From model upload to risk verdict.

01

Submit

Upload a model bundle via dashboard, API, CLI or CI/CD.

02

Select policy

Choose healthcare, finance or baseline — or a custom TOML profile.

03

Static scan

Six detection families run in parallel — no model execution.

04

Review

Detailed findings plus 15+ model-hygiene metrics.

05

Act

A clear Pass, Warn or Fail verdict gates the model.

/ Policy profiles

Tuned for regulated environments.

Apply different thresholds by sector, risk appetite and deployment context — all TOML-configurable.

Healthcare

For sensitive, PHI-adjacent workflows where model safety, auditability and vendor due diligence are critical.

Finance

Stricter controls on remote code loading, unsafe dependencies, model provenance and release approvals.

Global baseline

A strong default for general enterprise use, internal AI governance and standard MLOps controls.

Custom · TOML

Tune thresholds, allowlists and verdict logic without touching application code.

/ Sample scenario

Blocking a risky vendor model before deployment.

A financial-services team receives a fraud-detection model from a third-party vendor, packaged as a ZIP and submitted to Probexa before entering the registry.

The model is blocked, the vendor is asked to remediate, and the scan report is retained as evidence of supply-chain due diligence.
fraud-model.zip · finance profile
Suspicious pickle-based objects
Remote loader references
Typosquatted dependency
Archive path-traversal risk
Finance-profile violations
VerdictFAIL
/ Integrations & outputs

Built for modern AI engineering.

Run it where your teams already work, and surface findings in the tools they already use.

Web dashboard

Upload bundles, pick a profile, and view findings and risk metrics visually.

API & CLI

FastAPI endpoints and a CLI for portals, model registries and scripted pipelines.

GitHub Actions

A composite action for automated model scanning inside CI/CD.

SARIF & JSON

SARIF 2.1.0 for code scanning; JSON for SIEM, GRC and tickets. SBOM planned.

/ Where it fits

A security gate for your model supply chain.

Model registry intake

Scan every model before it enters your registry.

Vendor model review

Inspect third-party models before approval.

CI/CD security gate

Block unsafe models in PRs and releases.

MLOps governance

A repeatable check inside existing workflows.

Compliance evidence

Structured outputs for audit readiness.

High-risk AI review

Stricter profiles for regulated workloads.

/ FAQ

Questions, answered.

Does Probexa execute the model?

No. Probexa is static analysis — it scans model bundles without loading or running the model, so untrusted models never reach your runtime.

What risks does it detect?

Unsafe pickle usage, dangerous code indicators, TensorFlow unsafe ops, Keras custom layers, ONNX anomalies, Hugging Face remote-loader patterns, dependency typosquatting, archive path traversal and suspicious weight patterns.

Can Probexa run in CI/CD?

Yes — a GitHub Actions composite action plus SARIF 2.1.0 output surface findings directly in GitHub code scanning and other SARIF-compatible tools.

Does it support policy customization?

Yes. Policies are TOML-based with configurable thresholds and allowlists; healthcare, finance and global-baseline profiles ship built in.

Is Probexa a replacement for AI red teaming?

No. Probexa focuses on static model-artifact and supply-chain scanning. It complements red teaming, model evaluation and runtime monitoring — including Sera and SafeEval.

Is it only for regulated industries?

No — any organization using ML models benefits. Regulated sectors gain most from the stronger governance, auditability and vendor controls.

/ Get started

Scan your models before they reach production.

Run a pilot scan on an internal, vendor or open-source model and see the findings, metrics and verdict for yourself.