← Theory library
AI Security & Ops theory
Prompt injection defence, PII redaction, guardrails, tracing, and cost governance.
Practise in the Code Lab →Easy
- 1. The AI production stackSecurity, observability, memory, and automation are the four operational pillars.
- 2. Prompt injectionUntrusted text that reaches the model is executable instruction.
- 3. Jailbreaks and refusal robustnessLayer independent checks instead of trusting the model to refuse.
- 4. PII detection and redactionRedact on the way in, restore on the way out, log neither.
Medium
- 5. Guardrails as typed validatorsValidate structure and content, then repair or refuse — never ship unchecked output.
- 6. Tracing LLM applicationsOne trace per user request, with a span for every model and tool call.
- 7. Online quality evaluationScore live traffic continuously; offline suites cannot see real inputs.
- 8. Cost governance and rate limitingBudget per user, per feature, and per tenant — with hard stops.
- 9. Memory infrastructure: Mem0, Zep, Letta, Redis, PostgresChoose a store by access pattern, then define what is worth remembering.
- 10. Workflow automation: n8n, Temporal, Airflow, PrefectPick a runner by durability needs, and make every AI step idempotent.
Hard
- 11. Deploying prompt and model changes safelyVersion prompts like code and roll them out behind flags.
- 12. Incident response for AI systemsKnow how to contain a bad prompt, a leak, or a provider outage.
- 13. Governance, compliance, and auditabilityRecord what the system did, on whose behalf, with which data.
- 14. Capstone: production-ready AI featureShip one feature with every operational control wired and proven.