Privacy
Protecting the people whose data flows through the system.
Handling personal data lawfully and carefully in AI systems: minimising what is sent, redacting identifiers, controlling retention, and honouring access and deletion requirements.
Practical measures that work: redact before sending, keep raw data in your own store and send only what the task needs, and make deletion reach the vector index and any memory layer.
Minimisation is the intervention with the best ratio of effort to risk removed: if a value never leaves your network, it cannot appear in a log, a transcript or someone else’s context. Where the full value is genuinely needed, keep the mapping on your side and pass a token.
The identifier that was not needed
Classifying support tickets where the customer name is irrelevant to the category.
Redact before the call and the name never leaves. The classification is unchanged, the retention question mostly evaporates, and no downstream system has to be trusted with a record you did not have to share.
- docsData controls in the OpenAI platform ↗
Training-use and retention controls, including default abuse-monitoring retention and Zero Data Retention eligibility. Vendor-specific, and no substitute for your own privacy review.
- standardLLM01: Prompt injection ↗
Defines direct versus indirect injection and the downstream effects — data disclosure, unauthorised tool use, command execution — with least privilege and human approval as the recommended controls.
Data Governance
Rules about what data may be used, by whom, and retained how long.
PII Redaction · Anonymisation
Removing or masking personal identifiers before the model sees them.
Data Leakage
Information escaping where it should not.
Memory System
Storage that gives an assistant continuity across sessions.