Pickle in the Middle: Vertex AI Bucket-Squatting Bug Lets Attackers Swap Models

TL;DR Summary
A security flaw in Google Cloud Vertex AI's Python SDK allowed an attacker with their own Google Cloud project to hijack a victim's model upload by pre-creating a predictable staging bucket and swapping the uploaded model before Vertex AI reads it, enabling code execution when models saved with pickle/joblib are loaded. Google patched the issue in v1.144.0 (randomized bucket names) and v1.148.0 (ownership verification); upgrade to 1.148.0+ and explicitly set staging_bucket, plus audit all google-cloud-aiplatform usage across notebooks, CI jobs, and pipelines. No CVE has been assigned as of publication, and no exploitation in the wild was observed.
- Google Vertex AI SDK Flaw Let Attackers Hijack Model Uploads via Bucket Squatting The Hacker News
- Pickle in the Middle – Hijacking Vertex AI Model Uploads for Cross-Tenant RCE Unit 42
- Google’s Vertex AI SDK could allow RCE through bucket squatting csoonline.com
- Google Cloud Vertex AI Vulnerability Lets Attackers Take Over and Poison AI Models gbhackers.com
- Google Cloud Vertex AI SDK flaw allowed model hijacking and code execution | brief | SC Media SC Media
Reading Insights
Total Reads
0
Unique Readers
27
Time Saved
3 min
vs 4 min read
Condensed
84%
613 → 98 words
Want the full story? Read the original article
Read on The Hacker News