Safety and Security
Prompt injection, limiting the damage, secrets, PII, input and output guardrails, and the lethal trifecta.
5 notes · 12 min total
- 1Prompt InjectionDirect and indirect prompt injection, why indirect is the more dangerous one, and why delimiters and a better system prompt cannot fully stop it.3 min
- 2Limit the DamageDesign so a successful injection cannot do much: least-privilege tools, confirmation before risky actions, model output treated as untrusted, all as layers of…2 min
- 3SecretsWhy secrets never go in prompts, how prompt context leaks, and why you should assume the system prompt will leak.2 min
- 4PIIHow to protect personal data in an LLM app: send only what the task needs, redact or pseudonymize before the prompt, and never log raw PII.2 min
- 5Input and Output GuardrailsFilters on what goes into and comes out of an LLM app, why they are one layer and not a complete defense, and the lethal trifecta of private data, untrusted…3 min