Guardrails and safety
| # | File | Covers |
|---|---|---|
| 01 | Prompt injection | OWASP LLM #1 — why it can’t be sanitised away, and what actually works |
| 02 | Guardrails and output validation | structured output, groundedness checking, layering, fail open vs closed |
| 03 | PII, privacy and the EU AI Act | PII leak surfaces, EU AI Act as of August 2026, GDPR |
| 04 | Red teaming | indirect injection, denial of wallet, judging at the tool boundary |
| 05 | Model governance and responsible AI | model cards, risk tiering, meaningful oversight, change control |
The three positions
Prompt injection is not SQL injection. There’s no grammar separating instructions from data, so there’s no parameterisation. Defences are architectural: least privilege per context, server-side authorization, approval gates, egress control.
Indirect injection is the serious variant. The user is the victim, not the attacker. Any agent reading email, web pages or shared documents is executing text written by a third party.
The EU AI Act’s transparency duties are live now. June 2026 amendments delayed most high-risk obligations to December 2027 and August 2028, but Article 50, GPAI enforcement powers and the penalty regime all took effect 2 August 2026.