The Agent Skills Directory

EXTERNAL_DOWNLOADS / REMOTE_CODE_EXECUTION (MEDIUM): The skill frequently uses npx agentdb@latest for initialization and management. This command downloads and executes the latest version of the agentdb package from the public npm registry. Since the author and package are not on the trusted sources list, this pattern allows for unverified remote code execution in the user's environment.
COMMAND_EXECUTION / PERSISTENCE (MEDIUM): The skill provides instructions to modify the agent's persistent configuration via claude mcp add agentdb npx agentdb@latest mcp. This establishes a long-term persistence mechanism where the agent will automatically execute the external agentdb package in future sessions.
DYNAMIC_EXECUTION (MEDIUM): The create-plugin command generates and executes code based on templates (e.g., decision-transformer). This runtime generation of executable logic increases the attack surface if the templates or the generator itself are compromised.
INDIRECT_PROMPT_INJECTION (LOW): This memory management skill is a primary surface for indirect prompt injection. Data stored in the memory patterns could contain malicious instructions designed to be executed when the agent retrieves and synthesizes context.
Ingestion points: insertPattern, storeMemory, and storeFact APIs in SKILL.md.
Boundary markers: No explicit sanitization or delimiters (like XML tags or triple quotes) are demonstrated for stored memory content.
Capability inventory: The skill possesses file-system read/write capabilities (via database files), CLI command execution, and MCP server hosting.
Sanitization: No evidence of input validation or escaping for the pattern_data or content fields.

AgentDB Memory Patterns