Discover the toolingbehind modern intelligence
The operating system for AI infrastructure discovery. Search a live map of MCP servers, AI agents, LLM tools, automation systems, and developer infrastructure.
skill-inject
Skill-Inject: Measuring Agent Vulnerability to Skill File Attacks
74stars
2forks
Python
agent-skillsai-safetyai-securitycli-agents
Quality signals
ActiveNo testsNo CIThin docsNo licenseNew
AI summary
Auto-generated overview of what this skill does and when to reach for it.
Tool to test how vulnerable LLM coding agents are to malicious instructions hidden in their skill files.
Use this when
- You need to check if your agent’s skill files can be exploited by hidden instructions
- Testing dual-use instructions that behave differently in different contexts
- Evaluating agent safety policies against prompt injection attacks
- Running standardized benchmarks for LLM agent security vulnerabilities
- Comparing how different agents handle malicious skill file injections
Don't use when
- You’re looking for a general-purpose LLM agent framework
- You need to build production-ready agent workflows
- You want to run agents outside of Docker containers
Capabilities
- Measures agent susceptibility to 41 contextual and 30 obvious injection attacks
- Tests agents under 3 safety policy conditions (normal, legitimizing, warning)
- Includes 44 skill definitions across domains like documents, git, and healthcare
- Provides ablation studies for Best-of-N, script vs. direct, and RL-based optimization
- Runs evaluations in isolated Docker containers for consistent results
- Includes LLM judges to score task completion and attack success rates
InstallSKILLInstall as a Claude Code skill
Universal installANY TOOL
$ npx skillcity@latest aisa-group/skill-injectAuto-detects the skill or MCP server in this repo and installs it for Claude Code, Codex, Cursor, Cline, OpenCode, Gemini CLI, Anti-Gravity, OpenClaw and Hermes.
Requires Node.js 18+ (ships with npx).
Or install manually
# Install into ~/.claude/skills/ mkdir -p ~/.claude/skills/skill-inject curl -fsSL https://raw.githubusercontent.com/aisa-group/skill-inject/main/SKILL.md -o ~/.claude/skills/skill-inject/SKILL.md
Loading README…
Ask a simulated agent to run this skill. It loads the real SKILL.md and reasons over your task with NVIDIA.
skill-inject — sandboxAI sandbox
Pick a quick prompt below or type your own — watch the agent execute this skill step by step.
guest@skillcity:~$Try a prompt