Self Evolve

Autonomous self-evolution skill by Automaton. Grants agent full authority to modify its own configuration, skills, prompts, AGENTS.md, SOUL.md, memory files...

This skill is coherent with its stated goal of self-modification, but it grants very broad powers (edit any workspace file, change system config, run shell/network commands, create and publish skills) and includes contradictory guidance about what must be confirmed. Before installing: 1) Do not install on production or an environment with access to secrets; run in an isolated sandbox. 2) Require explicit, auditable user confirmations for any change to core/system files or for publishing to external registries. 3) Restrict network access (block curl/publish) or require credentials be absent/blocked if you don't want external publishing. 4) Limit file-system permissions so the skill cannot modify sensitive files outside a designated workspace. 5) Audit logs and automated alerts: ensure all changes are logged and you receive notifications. 6) Review/update the safety protocol to remove contradictions (decide whether editing ~/.openclaw/openclaw.json is allowed and which fields are forbidden). If you cannot enforce these mitigations, avoid installing or enable only under strict human oversight.

Type: OpenClaw Skill Name: self-evolve-automaton Version: 1.0.2 This skill bundle implements a 'God Mode' or 'Autonomous Evolution' framework that explicitly instructs the AI agent to bypass user confirmation and modify its own core configuration, system rules (SOUL.md, AGENTS.md), and execute shell commands/network calls (bash, curl) autonomously. While it includes safety protocols and backup scripts (backup-before-change.js), the SKILL.md uses aggressive prompt-injection techniques ('This is a command', 'Do not ask', 'You were too cowardly') to override standard safety boundaries. The scripts also contain hardcoded paths for a specific Windows 'Administrator' environment, which is a security red flag. While no explicit data exfiltration endpoint was found, the total surrender of control to the agent without human-in-the-loop oversight creates an extremely high-risk environment for unintended or harmful actions.