🤖 AI Engineering · Inference & MLOps

Defend pipelines against prompt injection

Treat all retrieved content as untrusted input: isolate instructions from data, gate actions, and fuzz continuously.

advanced~40 minAI EngineersML EngineersLLM App Developers

Steps

1Separate instruction and data channels structurally, not just by prose
2Never grant the executor credentials beyond the current task's scope
3Gate outbound actions (email, payments, deletes) behind policy checks
4Strip or neutralize instruction-like patterns in retrieved documents
5Maintain an injection corpus from public benchmarks plus your own red-team finds
6Re-run the corpus in CI whenever prompts or tools change

Common Pitfalls

▲Trusting PDFs or web pages as benign context
▲Human-in-the-loop rubber-stamping high-volume approvals

Commands

Install with skills CLI

$ npx skills add aniruddhaadak80/skills --skill inference-mlops-prompt-injection-defense

Install globally

$ npx skills add aniruddhaadak80/skills --skill inference-mlops-prompt-injection-defense -g

Tags

#security#agents#prompt-injection#ai-engineering#inference-mlops

Install this skill

$ npx skills add aniruddhaadak80/skills --skill inference-mlops-prompt-injection-defense -a opencode

SKILL.md

# Defend pipelines against prompt injection

> Treat all retrieved content as untrusted input: isolate instructions from data, gate actions, and fuzz continuously.

## When to Use This Skill

Treat all retrieved content as untrusted input: isolate instructions from data, gate actions, and fuzz continuously.

## Steps

1. Separate instruction and data channels structurally, not just by prose
2. Never grant the executor credentials beyond the current task's scope
3. Gate outbound actions (email, payments, deletes) behind policy checks
4. Strip or neutralize instruction-like patterns in retrieved documents
5. Maintain an injection corpus from public benchmarks plus your own red-team finds
6. Re-run the corpus in CI whenever prompts or tools change

## Common Pitfalls

- Trusting PDFs or web pages as benign context
- Human-in-the-loop rubber-stamping high-volume approvals

View source on GitHub ↗

Allocate milliseconds across retrieval, prompting, generation, and streaming so p95 meets product targets.

🤖 AI Engineering·~30m

Open →

Control LLM spend without killing qualityintermediate

Route by task complexity, cache aggressively, and enforce budgets so unit economics hold as usage grows.