Model Inversion & IP Exposure
Why ‘it trains on input’ is a confidentiality problem — and the three leak patterns.
The big idea
Model inversion is the risk that data fed into a model which trains on user input can be extracted later. That's exactly why the PSPF Advisory restricts use to providers with contractual no-training defaults (or a completed FOCI assessment). For contractors, three IP-leak patterns recur: a code leak (Samsung-style), a document leak (NSW RA-style), and a capability leak — drafting tenders or white papers with capability-revealing content in a public LLM that retains and may train on it.
Key insight: Assume anything placed into a consumer AI that trains on input may resurface later. Enterprise/API services with contractual no-training guarantees are the baseline for anything you'd mind seeing again.
What you'll see
Model inversion and the three contractor IP-leak patterns.
The risk and the patterns
What to guard against. Tap to expand.