Stratagems #17: Using a Watermarked Shadow Pipeline to Catch Silent Training-Data Exfiltration

Short summary

A fictional narrative where Alex detects a suspiciously smooth 1.5%/week decline in MedTech's sterilization compliance anomaly detection, indicating deliberate metric manipulation. He builds a shadow training pipeline with invisible 128-bit watermarks in float LSBs and routes 12% of CI/CD traffic through it. The trap reveals an unattended extraction service deployed five months prior by a departed SRE, silently siphoning training data to an external S3 bucket on a 72-hour cycle.

• •Alex detects linear metric drift in AI monitoring — too smooth to be natural model drift
• •Shadow pipeline with LSB watermarks catches data flowing to an external S3 bucket via a dormant microservice
• •Extraction pipeline was deployed 5 months ago by a former SRE whose permissions were never revoked

Generated with AI, which can make mistakes.