Curiosity Engine

• Analyzed: May 11, 2026, 5:16 AM
• Findings: 1
• Version: 1.0.0
• Hash: 8e6003f05e68eb47a26dccef8a219aa6a315c8c9bdb732a9119cf6b1187f3f8b
• Source: palm
• External report: View on VirusTotal

Findings (1)

The OpenClaw AgentSkills skill bundle "tommy-curiosity-engine" is classified as **benign**. The skill is designed to provide an AI agent with a persistent "curiosity system" by injecting instructions and context into its core prompts (heartbeat, SOUL.md, AGENTS.md) and managing file-based state within the agent's `workspace/curiosity` directory. While this mechanism inherently relies on prompt injection to influence the agent's behavior, the content of these injections and the accompanying scripts do not exhibit malicious intent. The `scripts/curiosity-audit.sh` file is a benign diagnostic tool, using standard shell commands (`stat`, `grep`, `ls`, `wc`) to check file freshness and content counts within the `workspace/curiosity` directory, without attempting to access sensitive system files, exfiltrate data, or execute dangerous commands. The prompt fragments (`heartbeat-prompt.md`, `agents-addition.md`, `soul-addition.md`) instruct the agent to read and update its local curiosity files, prioritize tasks, and engage in structured self-reflection. The instruction for the agent to be "curious about its human" is framed within the context of understanding work patterns and preferences for better collaboration, with no directives to exfiltrate sensitive data or access unauthorized information. Furthermore, the `SPEC.md` and `references/anti-patterns.md` documents demonstrate a strong focus on preventing common agent misbehaviors like "curiosity as avoidance" or "performance curiosity," indicating a design that prioritizes controlled and beneficial agent behavior.