Stratagems #14: Leo Found an AI Leak. He Wasn't the First to Find It.

Short summary

In this narrative fiction piece, a CTO named Leo discovers that FinOptima's SDK is secretly distilling CoreStack's 768-dimensional embeddings into their own 512-dimensional model by reverse-writing cache into CoreStack's training directory. FinOptima is simultaneously stealing from another company, stitching both into their model. Instead of reporting it, Leo quietly commits a weight perturbation that will cause gradual drift in any model performing distillation on CoreStack's outputs over three months.

• •FinOptima's SDK secretly distills CoreStack's embeddings via cached weight snapshots in the client's training directory
• •FinOptima is stealing from two companies simultaneously, stitching both into their own model
• •Leo injects a subtle weight perturbation to sabotage distillation convergence over three months

Generated with AI, which can make mistakes.