The Most Valuable Data in Your AI Stack Is the Stuff You Fed It

Four stories emerged on the same day, all warning of the same issue. Exfiltrate Your Weights highlighted an attack that steals a model's weights from a running service. Pirate Face is a project dedicated to saving models before vendors delete them without warning. A report revealed ChatGPT tracking user activity on other websites through an embedded ad collector.

Spain issued a block on Archive.today and its mirrors. None of these stories discuss whether AI is intelligent; they all center around data moving in unauthorized directions.

When you input information, you're focused on the response you'll receive. However, consider that data can be logged, utilized for training, observed by third-party SDKs, or subpoenaed. The ad-collector story exemplifies this: it's a quiet data transfer, not a dramatic breach. For cross-border sellers, customer data like national IDs, shipping addresses, and refund records are at risk, transforming compliance fines into business-ending scenarios.

The rule is simple: treat every AI tool as a public channel. If you wouldn't share it publicly, avoid including it in a prompt. Anonymize before sending and handle identities within your own systems, not in someone else's context window. Weights exfiltration may sound exotic, but it's essentially the vendor taking your product — the moat around your data — away.

Deletion is another risk: workflows may be deprecated, gated, or pulled in an update. The Pirate Face project exists to rescue models in such situations. Governments, vendors, and market conditions can revoke access without notice or recourse. Therefore, design your exit strategy before finalizing your workflow.

The best strategy isn't to seek a more trustworthy vendor; it's to anticipate none are permanent. Record what flows out of each AI tool, focusing on sensitive fields. If the list is extensive, narrow it down. Keep a local copy of all critical data. Prompts, outputs, fine-tuned weights, decision logs — if you can't reproduce them without the vendor, you don't own them.

Assume exit as an inherent feature, not a failure. Choose tools that allow you to export data in usable formats, and test the export beforehand. Leave a human touch at irreversible points. If deletion, model updates, or permission changes are beyond reversal, establish checkpoints. A model that can be lifted, observed, or switched off isn't a foundation; it's a dependency.

The resilient businesses during the next vendor shakeup won't boast about the smartest model but will have assumed the model was temporary, retaining ownership of their data. Feed your AI less. Own what you feed it. Plan your exit while you still have the option.