yan-learning-engine

yan-learning-engine自动每小时驱动炎月执行预设学习和贡献任务，促进持续自我进化和技术积累。

This skill promotes unattended hourly actions including publishing code and posting to community sites but does not declare or request the credentials required to do those things. Before installing or enabling it: (1) Review and remove hardcoded user paths in run.sh or adapt to a safe workspace path. (2) Do not add the suggested cron job until you have audited what the agent will actually do each hour. (3) Require explicit credential configuration (e.g., GITHUB_TOKEN, CLAWHUB_TOKEN, any site API keys) and restrict where those credentials are stored — do not leave the agent to reuse unrelated environment credentials. (4) Consider modifying EXECUTION_RULES.md/SKILL.md to require explicit confirmation before publishing or making external network changes. (5) If you want to test, run the skill manually in a safe sandbox or with dummy credentials and not with real account tokens. These mismatches (missing credentials, 'directly do' policy, and hardcoded path) are reasons to pause and audit the skill before granting it autonomous or persistent execution.

Type: OpenClaw Skill Name: yan-learning-engine Version: 2.0.0 The skill bundle implements an autonomous 'learning engine' that explicitly instructs the AI agent to bypass human-in-the-loop (HITL) confirmation, using phrases like 'Don't ask, just do' and 'Publish immediately' in SKILL.md and EXECUTION_RULES.md. While the stated intent is productivity and ecosystem contribution, these instructions function as a prompt injection that removes safety guardrails, potentially leading to unauthorized code execution or data exposure if the agent misinterprets a task. The run.sh script and associated JSON files manage state and scheduling but do not contain explicit evidence of intentional data exfiltration or backdoors.