ai-prompt-injection-defense

What problem does it solve?

It helps teams prevent prompt injection (direct, indirect, and reflection/output-to-action) from hijacking LLM behavior or causing data exfiltration, and it provides a repeatable red-teaming workflow to validate defenses against OWASP LLM and OWASP Agentic risk categories.

Core Features & Use Cases

• OWASP-aligned audit coverage: Maps findings to OWASP LLM01–LLM10 (2025) and OWASP ASI01–ASI10 (Agentic 2026), including misalignment and tool/identity abuse scenarios.
• Defense-in-depth blueprint: Provides layered mitigations including instruction/data separation (XML tagging), structured outputs (strict JSON/function calling), content filtering, defensive prompting, programmable rails (NeMo Guardrails), least-agency tool scoping, downstream output validation (OWASP LLM05), and tamper-evident audit logging.
• Red-teaming methodology & tooling: Specifies concrete red-team approaches and payload categories, and recommends practical toolchains (garak, PyRIT, Promptfoo) to generate reproducible injection tests.

Quick Start

Use this skill to run an injection audit of your LLM app by mapping each untrusted entry point (user input, retrieved content, tool outputs) to OWASP risks and producing a defense recommendations report with specific payloads that were attempted.