We commit to build and communicate privacy processes that protect the data of stakeholders interacting with our AI systems directly and indirectly, enabling Trust by Privacy.
AI systems widen the paths along which personal data can leak. Models can memorise training data and reproduce it under extraction attacks, a builder-side risk that fine-tuners take on the moment they train on their own records. On the deployer side, every inference call can send user data to third parties the person never agreed to, and shared agent infrastructure can let one user’s data surface in another’s session.
Privacy should therefore be designed at every layer: assessed before data is connected to a model, segregated across users and tiers in memory, and encrypted in transit and at rest. Trust also requires communication, so stakeholders, direct and indirect, should be able to learn what data is held, how it is handled and why protecting it matters.
01 — WHERE IT FAILS
Where it fails
The leak paths are wider than the application’s own database, and most of them sit outside classic privacy reviews.
- Personal data extractable from a model by prompting.
- Sensitive data flowing to third parties the person never agreed to.
- One user’s data surfacing in another user’s session.
02 — PRACTICAL CONTROLS
Practical controls
Controls should protect data at every layer it passes through, before it reaches a model, while it sits in memory, and wherever it travels.
- Run a privacy impact assessment before connecting data to a model.
- Segregate memory across users, with no shared session context.
- Encrypt personal data end to end, in transit and at rest.