The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions (in SKILL.md) direct the agent to pass data to a local Python script using a shell pipe (e.g., echo '...' | python3 scripts/graph_check.py). This pattern introduces a potential command injection vulnerability if the data retrieved from memory contains shell metacharacters like single quotes and is not correctly escaped by the agent before the shell command is constructed.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes structured data from an external memory source and lacks explicit safety instructions to prevent the execution of embedded commands.
Ingestion points: The agent retrieves entity and relation data via the sage_memory_search tool as described in SKILL.md.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are present in the data processing flow.
Capability inventory: The skill utilizes storage management tools (sage_memory_store, sage_memory_delete, sage_memory_update) and executes a local validation script via subprocess (file: SKILL.md).
Sanitization: The instructions do not provide guidance on sanitizing or escaping the data before it is piped to the shell or processed by the validator.
[SAFE]: The included validation script (scripts/graph_check.py) was audited and found to be safe. It relies only on standard Python libraries, performs no file system or network operations, and exclusively processes input through standard input.
[SAFE]: The skill explicitly warns against storing sensitive credentials such as passwords, tokens, or API keys within the knowledge graph, demonstrating a proactive approach to data security.

ontology