Mupeng Evolve — God-Tier Agent Evolution Engine

The God-Tier Agent Evolution Engine. 6 top skills analyzed, dissected, and rebuilt from scratch. Zero external dependencies. Battle-tested in production. Mak...

This skill is instruction-only and claims 'zero dependencies', but the runtime instructions expect command-line tools (git, find) and allow the agent to modify other skills and write to repo files. Before installing: 1) Ask the author to clarify and list required binaries and exact file paths the agent will read/write. 2) Require a safe, auditable human-approval enforcement mechanism for any edits to core files or other skills; don't rely on a natural-language 'security gate' without enforcement. 3) Run the skill in a restricted sandbox or test workspace (back up your skills and repo first). 4) If you enable autonomous invocation, restrict its permissions (prevent writes to other skills or system-wide config) until you have audited the exact mutation logic. 5) Request detail about the 'Revenue Flywheel' section — confirm whether any external API calls or credentials are required and insist those be declared. If the author cannot clearly justify why the skill needs to edit other skills or cannot enumerate concrete enforcement of the human-approval gates, treat it as high-risk and avoid enabling autonomous modifications.

Type: OpenClaw Skill Name: mupeng-evolve Version: 1.0.0 The skill bundle is classified as suspicious due to the broad and powerful capabilities it instructs the AI agent to perform, which, despite stated security guardrails, present significant attack surface and potential for misuse. Specifically, the `SKILL.md` instructs the agent to 'evolve freely' by modifying its own `tools` and `skills`, and to extensively interact with the file system using commands like `git log` and `find`. While the skill explicitly states 'Absolute block: secrets, API keys, auth tokens — never touched' and includes a 'SECURITY SCAN' for 'Injection detected,' the ability for an agent to self-modify its operational code and configuration, coupled with implied external interactions for 'auto-publish' and 'sub-agent execution,' introduces high-risk vulnerabilities if not perfectly sandboxed or if the agent's interpretation deviates from the intended benign purpose. There is no clear evidence of intentional malicious behavior, but the inherent risks of these capabilities warrant a 'suspicious' classification.