Skip to content

AI/ML Security with JFrog Xray

← Back to JFrog AI & ML

The AI/ML ecosystem has a significant and growing software supply chain risk. Popular libraries like PyTorch, TensorFlow, Transformers, and LangChain pull in hundreds of transitive dependencies — any of which could carry critical CVEs. JFrog Xray provides automated scanning across your entire artifact repository.

All steps use JFrog SaaS at https://<company>.jfrog.io. Xray is included in JFrog SaaS plans.

What Xray Scans for AI/ML Projects

Scan Type What It Detects
SCA (Software Composition Analysis) CVEs in Python packages (PyPI), Conda packages
License Compliance Packages with restrictive licenses (GPL, AGPL) in commercial products
Operational Risk Deprecated, unmaintained, or low-quality packages
Secrets Detection Hard-coded credentials in code or model configs
SAST Static application security testing in ML training scripts

Step 1: Enable Xray Indexing on ML Repositories

By default Xray scans all repositories. Verify your ML repos are indexed:

  1. Go to Administration → Xray → Indexed Repositories
  2. Confirm these repos are listed and active:
  3. pypi-local
  4. pypi-remote
  5. ml-models-local
  6. ml-models-staging-local
  7. If not, click + Add Repository and select them

Step 2: Create a Security Policy

A Security Policy defines what constitutes a violation — for example, any CVE with CVSS score ≥ 7.

  1. Go to Administration → Xray → Policies
  2. Click + New Policy
  3. Set Policy Name: ml-security-policy
  4. Set Policy Type: Security
  5. Under Rules, click + New Rule:
  6. Rule Name: block-critical-cves
  7. Criteria: Min Severity = Critical (CVSS ≥ 9.0)
  8. Automatic Actions:
    • ✅ Block Download (prevent artifact from being pulled)
    • ✅ Fail Build (fail the CI build)
    • ✅ Notify (send email/Slack alert)
  9. Click Save

Step 3: Create a Watch

A Watch connects repositories to policies — it defines what to watch and which policy to apply.

  1. Go to Administration → Xray → Watches
  2. Click + New Watch
  3. Set Watch Name: ml-repos-watch
  4. Under Resources, add:
  5. pypi-local, pypi-remote, ml-models-staging-local
  6. Under Assigned Policies, add ml-security-policy
  7. Click Save

Now all new packages indexed in those repos will be automatically checked against the policy.

Step 4: Scan a Build in CI/CD

After running your training pipeline and publishing build info, scan it:

# Scan the specific ML training build
jf rt build-scan my-ml-model 42

# Returns exit code 1 if Xray violations found (fails CI)

CI output example:

[INFO] Scanning build my-ml-model #42...
[ERROR] Found 1 policy violation:
  CRITICAL CVE-2023-29483 in requests:2.28.0
  Description: SSRF vulnerability in requests library
  Blocking download: Yes
[INFO] Build scan complete. Exit code: 1

Step 5: View Xray Scan Results

  1. Navigate to Application → Security & Compliance → Scans
  2. Select Build: my-ml-model
  3. Switch to the Security tab

You'll see:

Package CVE Severity Fixed Version
requests:2.28.0 CVE-2023-29483 Critical 2.31.0
pillow:9.2.0 CVE-2023-44271 High 10.0.0

Step 6: Review AI Package Risks

Common AI/ML packages with security histories:

Package Risk Concern
langchain Rapid release cycle; supply chain injection risks
pytorch Large attack surface; transitive dependency CVEs
transformers Pickle deserialization risk in model loading
tensorflow Historical buffer overflow and memory corruption CVEs
numpy Older versions with integer overflow issues

Xray's Contextual Analysis (JFrog Advanced Security) can determine if a CVE is actually exploitable in your code — reducing false positives.

Step 7: Enforce Governance with Fail-Build Actions

In your JFrog policy, enable "Fail Build" to enforce that only clean builds get promoted:

# In CI: this fails if any Critical/High CVEs found
jf rt build-scan my-ml-model ${BUILD_NUMBER} || exit 1

# Only promote if scan passed
jf rt build-promote my-ml-model ${BUILD_NUMBER} \
  --source-repo ml-models-staging-local \
  --target-repo ml-models-prod-local

Use Cases

Scenario Solution
PyTorch update has a CVE Xray flags it; Curation blocks download from pypi-remote
ML training script uses pickle.load on untrusted data JAS SAST flags deserialization risk
GPL library embedded in commercial ML product License policy violation raised
requests library outdated in training environment SCA scan reports old version with known fix
Security audit requires SBOM Xray generates SBOM for any build

Next Steps

👉 Curating AI/ML Packages 👉 MLOps Pipeline with JFrog

🧠 Quick Quiz

#

In JFrog Xray, what component connects a repository to a security policy to trigger automated scanning?

📬 DevopsPilot Weekly — Learn DevOps, Cloud & Gen AI the simple way.
👉 Subscribe here