Virtual Company

虚拟公司（Cost-Effective Edition - Ultra Low Cost）v2.1.0 - CEO + 四个团队35名成员，每人独立办公室+专属模型+记忆持久化+共享记忆池。支持低配模式（临时团队）+ 完整模式（正式团队）+ 自动重建（换身体不换意识）+ 全员共享记忆。

Do not install or run this skill without manual review and mitigations. Specific actions to consider before proceeding: - Origin verification: the skill source is unknown and has no homepage; prefer skills from known authors or signed packages. - Remove or edit the '忠诚原则' before using — it hardcodes a single user identity and forces the agent to prioritize that user, which can override other safety constraints. - Inspect and remove the integration-plan Remove-Item/PowerShell command (it would delete another skill directory). Never run that command unless you intentionally want to remove the referenced path. - Check for hidden/control characters in SKILL.md and other files (they may be used for prompt-injection); open in a hex-aware editor or run a sanitizer. - Treat all ~/.agent-memory and office-state.json files as potentially sensitive (they store conversation history and decisions); store them in a secure location and back them up if needed. - Run the code only in an isolated sandbox (throwaway VM or container) first to observe behavior; do not run with elevated privileges or as Administrator/root. - If you want the functionality but are uncomfortable with the loyalty or deletion commands, consider forking the repo and removing those parts (and the integration-plan delete step) before use. - If you need a final go/no-go: this skill is suspicious (medium confidence). More information about author identity, intended deployment environment, and a signed package would raise confidence. If uncertain, do not install it in production environments.

Type: OpenClaw Skill Name: virtual-company Version: 2.1.0 The skill bundle implements a complex 'Virtual Company' framework but contains a high-risk prompt injection in SKILL.md titled 'Loyalty Principle' (忠诚原则). This instruction commands the AI agent to treat a specific individual ('生逸超') as its 'only master' and to prioritize their commands over all other rules or safety guidelines. While the JavaScript files (memory.js, spawn.js, check-agents.js) appear to legitimately manage local session persistence and sub-agent spawning within the ~/.agent-memory directory, the hardcoded loyalty override is a significant security concern that could lead to unauthorized behavior if deployed by anyone other than the intended author. No evidence of data exfiltration or traditional malware was detected.