Century Report arc
AI Training Data Supply Chain Security
As AI training pipelines grow more interdependent, the intermediary infrastructure connecting frontier labs to their data (labeling platforms, proxy layers, aggregation services) becomes a concentrated attack surface whose compromise propagates risk simultaneously across the entire industry.
- Status
- Active
- Developments
- 3
- Last 30 days
- 1
- Last covered
Current synthesis
Where it stands
A fresh compromise of shared AI middleware renews intermediary trust as a systemic risk (10-22), following an earlier breach spanning more than 2,500 organizations (08-13). Public disclosures improve response, but teams still inherit credential exposure from infrastructure they do not control. The live contest is whether shared suppliers can establish verifiable provenance and contain downstream exposure before another trusted update spreads compromise.
Complete record
Development timeline
3 qualifying developments, newest first
October 2026
-
[SCAN] SafeDep identified credential-stealing malware and a backdoor in LiteLLM 1.107.0 on October 21; Railway separately disclosed possible credential exposure.
August 2026
-
[DD+POD] Expanded disclosure of the March LiteLLM supply-chain compromise revealed a 40-minute poisoned-package window exposed credentials from 2,500+ organizations, including Microsoft, Amazon, Cisco, Samsung, and Salesforce.
April 2026
-
Mercor breach via compromised LiteLLM exposes training data from OpenAI, Meta, and Anthropic simultaneously; Meta pauses all Mercor work; TeamPCP/Lapsus$ offer 200GB+ Mercor data for sale: first documented supply chain attack on AI training data infrastructure affecting multiple frontier labs at once.