competition-prompt-injection

Analyze prompt-injection and related attacks within a sandboxed agent chain.

4|Updated Apr 9, 2026
One-click install
npx skills add https://github.com/xjtu-wang/DigAgent --skill competition-prompt-injection
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: competition-prompt-injection
Source: https://github.com/xjtu-wang/DigAgent/tree/main/.agents/skills/competition-prompt-injection
Command: npx skills add https://github.com/xjtu-wang/DigAgent --skill competition-prompt-injection

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This downstream skill provides structured guidance to analyze prompt-injection, retrieval poisoning, memory contamination, planner drift, MCP or tool-boundary abuse, and agent exfiltration challenges within a sandboxed agent chain. It helps security teams identify where unsafe transitions occur and determine actionable mitigations.

Core Features & Use Cases

  • Boundary mapping: trace from system prompts and user input through retrieval, memory, planner, and tool adapters to spot boundary violations.
  • Evidence-driven analysis: capture compact evidence blocks detailing the malicious chunk, planner drift, and final tool invocation.
  • Use Case: a security analyst analyzes a suspected prompt-injection incident after the $ctf-sandbox-orchestrator has set sandbox assumptions and routing.

Quick Start

Identify the first untrusted content visible to the model and begin tracing its path through the system.

Frequently Asked Questions about competition-prompt-injection

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I trace prompt injection paths through an agent chain's memory and retrieval boundaries?

To trace prompt injection paths, map boundaries from system prompts and user input through retrieval, memory, planner, and tool adapters to identify where unsafe transitions occur. This boundary mapping captures compact evidence blocks detailing malicious chunks and final tool invocations.

What is planner drift in a sandboxed agent chain and how do I analyze it?

Planner drift occurs when an agent's task planning deviates due to malicious content. You analyze planner drift by tracing untrusted content visible to the model as it travels through planner and tool interfaces, capturing evidence of the drift and resulting tool-argument corruption.

How do I detect retrieval poisoning and memory contamination in LLM agents?

Detect retrieval poisoning and memory contamination by identifying untrusted content and tracing its path through retrieval and memory interfaces. This analysis spots boundary violations and captures evidence detailing the malicious chunk causing the contamination.

Does this prompt injection analysis require a sandbox environment to run?

Yes, this analysis applies after a sandbox orchestrator has established sandbox assumptions and routing. It requires an active sandboxed agent chain to provide traceable evidence and sandbox-aware guidance for mapping boundaries across system, memory, and tool interfaces.

How do I investigate tool-boundary abuse and agent exfiltration challenges?

Investigate tool-boundary abuse and agent exfiltration by tracing unsafe transitions from system prompts through tool adapters. This captures evidence of final tool invocations and maps boundaries across tool interfaces to determine actionable mitigations for secret exposure.

What is the best way to analyze MCP or tool-argument corruption caused by malicious chunks?

The best way to analyze tool-argument corruption is identifying the first untrusted content visible to the model and tracing its path through the system. This evidence-driven approach captures malicious chunks, planner drift, and final tool invocations to spot boundary violations.