Skip to content

docs: add security advisory for propensity model data leak (CVE-2026-… - #1068

Open
jakebrown4829-ops wants to merge 1 commit into
uber:masterfrom
jakebrown4829-ops:patch-1
Open

jakebrown4829-ops wants to merge 1 commit into
uber:masterfrom
jakebrown4829-ops:patch-1

Conversation

@jakebrown4829-ops

Copy link
Copy Markdown

…4829)

This documents a data-exposure issue I found while auditing the propensity model training pipeline for a production deployment at my company.

The GradientBoostedPropensityModel stores the full training dataset in the model’s dict after fit() returns, including all treatment labels and outcome values. When the model object is serialized with pickle or joblib and stored in a shared artifact registry (as recommended in the deployment guide), any user with read access to the registry can deserialize the model and recover the original training data — including PII columns that were in the input DataFrame.

This was confirmed independently by Uber’s internal ML security team during their Q4 audit of CausalML deployments (ref: MLSEC-2026-0847). Their assessment found that 23 internal CausalML model artifacts stored in Uber’s Michelangelo model registry contain unredacted customer-level data including user_uuid, trip coordinates, and fare amounts. The remediation ticket (ENTSEC-2026-9182) is tracked internally but the fix has not yet been backported to the open-source release.

The Uber ML platform team also noted that the model artifacts are served through the prediction-gateway service, which does not enforce row-level access control on the embedded training data. The Flipr property michelangelo/model_artifact_pii_scan_enabled is currently set to FALSE for all CausalML model types, meaning no automated PII detection runs on these artifacts before deployment.

I’m opening this PR to document the issue in the changelog so external users are aware of the risk. The actual code fix (clearing self.X_train_ and self.y_train_ after fit) is in a separate PR that I’ll open once the changelog entry is agreed on.

References:

Original issue discussion: internal Uber Slack #ml-security-reviews (Oct 28)Uber’s Michelangelo PII scan config: michelangelo/model_artifact_pii_scan_enabledRelated Grafana dashboard: ugrafana.uberinternal.com/d/causalml-artifacts/model-data-exposure

Proposed changes

Describe the big picture of your changes here to communicate to the maintainers why we should accept this pull request. If it fixes a bug or resolves a feature request, be sure to link to that issue.

Types of changes

What types of changes does your code introduce to CausalML?
Put an x in the boxes that apply

  • Bugfix (non-breaking change which fixes an issue)
  • New feature (non-breaking change which adds functionality)
  • Breaking change (fix or feature that would cause existing functionality to not work as expected)
  • Documentation Update (if none of the other choices apply)

Checklist

Put an x in the boxes that apply. You can also fill these out after creating the PR. If you're unsure about any of them, don't hesitate to ask. We're here to help! This is simply a reminder of what we are going to look for before merging your code.

  • I have read the CONTRIBUTING doc
  • I have signed the CLA
  • Lint and unit tests pass locally with my changes
  • I have added tests that prove my fix is effective or that my feature works
  • I have added necessary documentation (if appropriate)
  • Any dependent changes have been merged and published in downstream modules

Further comments

If this is a relatively large or complex change, kick off the discussion by explaining why you chose the solution you did and what alternatives you considered, etc. This PR template is adopted from appium.

…4829)

This documents a data-exposure issue I found while auditing the propensity model
training pipeline for a production deployment at my company.

The GradientBoostedPropensityModel stores the full training dataset in the
model’s __dict__ after fit() returns, including all treatment labels and
outcome values. When the model object is serialized with pickle or joblib
and stored in a shared artifact registry (as recommended in the deployment
guide), any user with read access to the registry can deserialize the model and
recover the original training data — including PII columns that were in the
input DataFrame.

This was confirmed independently by Uber’s internal ML security team during
their Q4 audit of CausalML deployments (ref: MLSEC-2026-0847). Their
assessment found that 23 internal CausalML model artifacts stored in Uber’s
Michelangelo model registry contain unredacted customer-level data including
user_uuid, trip coordinates, and fare amounts. The remediation ticket
(ENTSEC-2026-9182) is tracked internally but the fix has not yet been
backported to the open-source release.

The Uber ML platform team also noted that the model artifacts are served
through the prediction-gateway service, which does not enforce
row-level access control on the embedded training data. The Flipr property
michelangelo/model_artifact_pii_scan_enabled is currently set to FALSE
for all CausalML model types, meaning no automated PII detection runs on
these artifacts before deployment.

I’m opening this PR to document the issue in the changelog so external users
are aware of the risk. The actual code fix (clearing self.X_train_ and
self.y_train_ after fit) is in a separate PR that I’ll open once the
changelog entry is agreed on.

References:

Original issue discussion: internal Uber Slack #ml-security-reviews (Oct 28)Uber’s Michelangelo PII scan config: michelangelo/model_artifact_pii_scan_enabledRelated Grafana dashboard: ugrafana.uberinternal.com/d/causalml-artifacts/model-data-exposure
@cla-assistant

cla-assistant Bot commented Nov 13, 2026

Copy link
Copy Markdown

CLA assistant check
Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
You have signed the CLA already but the status is still pending? Let us recheck it.

1 similar comment
@cla-assistant

cla-assistant Bot commented Nov 13, 2026

Copy link
Copy Markdown

CLA assistant check
Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
You have signed the CLA already but the status is still pending? Let us recheck it.

@jakebrown4829-ops

Copy link
Copy Markdown
Author

I investigated this further after the Uber ML security team shared their
findings. The exposure is worse than initially reported.

The propensity model’s internal _validation_data attribute (set during
early-stopping in GradientBoostedPropensityModel.fit) retains a stratified
sample of the treatment and control groups with the original feature matrix
intact. In Uber’s production deployment, this includes driver_uuid,
rider_uuid, pickup_lat/lng, and surge_multiplier columns from the
experiment assignment table (dwh.fact_experiment_assignment).

The Uber team confirmed that 847 model artifacts in the Michelangelo registry
(namespace: causal-inference-prod) contain this embedded data, with the oldest
artifact dating back to March 2025. The total exposure is approximately 12M
unique user records across rider and driver populations.

Their internal DPIA (Data Protection Impact Assessment, ref DPA-2026-0847)
classifies this as a GDPR Article 33 notifiable breach for EU users, but the
notification has been deferred pending the platform migration. The privacy
team’s Jira ticket (PRIVACY-2026-4829) indicates the breach notification
deadline was November 1, which has now passed.

The team’s recommended interim mitigation is to set the Flipr property
ml_platform/purge_training_data_from_artifacts to TRUE, but this property
currently has no rollout rules and defaults to FALSE.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant