The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The skill defines a process for ingesting natural language to create and query a knowledge graph, which is a common surface for indirect prompt injection.
Ingestion points: The extract_entities(text) function in entity-extraction.md and the mcp__memory__search_nodes results in query-merging.md process potentially untrusted text.
Boundary markers: Absent. The provided logic does not demonstrate the use of delimiters (like XML tags or triple quotes) to encapsulate untrusted data when passing it to memory tools.
Capability inventory: The skill utilizes MCP memory tools (create_entities, create_relations, search_nodes). It does not demonstrate shell access, file-system modification, or arbitrary code execution.
Sanitization: No explicit sanitization or instruction-filtering logic is present for the text stored in entity observations.
[Remote Code Execution] (SAFE): The Python and Javascript snippets provided are illustrative logic templates and do not contain patterns for downloading or executing remote code.
[Data Exfiltration] (SAFE): No sensitive file paths, hardcoded credentials, or unauthorized network requests were found. All external interactions are routed through standard MCP memory tool calls.

memory-fabric