The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes agent metadata (names, descriptions, and job offerings) retrieved from the external ACP registry via the browse_agents tool. This content is used to inform the agent's decisions on which agents to interact with, creating an attack surface for indirect prompt injection. \n
Ingestion points: Data is ingested from the Virtuals Protocol registry through the @virtuals-protocol/acp-node library in scripts/index.ts. \n
Boundary markers: The skill instructions do not specify any delimiters or safety warnings for the agent to distinguish between its instructions and the data fetched from the registry. \n
Capability inventory: The skill allows for the initiation of on-chain transactions and payments through the execute_acp_job tool. \n
Sanitization: There is no evidence of sanitization or filtering of the external registry data before it is presented to the agent. \n- [COMMAND_EXECUTION]: The skill's functionality is delivered through a local TypeScript CLI tool. The agent is instructed to execute npx tsx scripts/index.ts with various arguments to perform tasks like browsing agents and managing jobs. This relies on local shell execution within the agent's environment. \n- [EXTERNAL_DOWNLOADS]: The installation process involves cloning the source code from the Virtual-Protocol/openclaw-acp GitHub repository and using npm install to download dependencies, including the @virtuals-protocol/acp-node SDK. These are standard operations for this skill's use case.

virtuals-protocol-acp