[NEW SKILL] multi-agent-prompt-injection-defense: defend against prompt injection across agent communication channels

Skill name

multi-agent-prompt-injection-defense

Domain / Subdomain

cybersecurity / adversarial-machine-learning

Description

Detect and defend against prompt injection attacks targeting AI agents in multi-agent orchestration systems. Covers poisoned context injection, inter-agent message tampering, and output filtering for agents that aggregate LLM outputs.

MITRE ATT&CK techniques

BobbyTables M07 — External Analysis, or request new technique assignment for prompt injection.

NIST CSF function

Detect (DE) and Respond (RS)

Why is this skill needed?

Multi-agent pipelines (e.g., LangChain, AutoGen, CrewAI, or custom orchestrators) route outputs from one agent directly into another agent's context. An attacker who poisons the output of one agent can inject instructions that subvert downstream agents, escalate privileges, or exfiltrate data. Existing skills in this repo cover single-agent prompt injection (detecting-ai-model-prompt-injection-attacks) but not the cross-agent variant. This skill would help security practitioners and AI engineers harden multi-agent systems against this attack class.