The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill includes an execute_blender_code tool in SKILL.md that allows the agent to run arbitrary Python code within the Blender application environment. This capability can be exploited if the agent is manipulated into executing malicious logic.
[EXTERNAL_DOWNLOADS]: Installation instructions in quickstart.md direct users to download and install a script from an external third-party repository (github.com/ahujasid/blender-mcp). Additionally, SKILL.md specifies the use of uvx to download and run the blender-mcp community package, which is not verified.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process network discovery data (CDP/LLDP) as described in SKILL.md. This data may contain untrusted strings, such as device hostnames, which could influence the agent's behavior when generating Python commands. Evidence: 1. Ingestion: CDP/LLDP neighbor data parsed from integrated skills (referenced in SKILL.md). 2. Boundary markers: Absent; no safety instructions or delimiters are used to wrap external data. 3. Capability inventory: execute_blender_code (arbitrary Python exec) and file-write capabilities for image export in SKILL.md. 4. Sanitization: Absent; hostnames are classified for coloring but are not sanitized before being used in tool operations.
[DYNAMIC_EXECUTION]: The execute_blender_code tool in SKILL.md enables runtime execution of dynamically generated Python scripts, making it difficult to verify the security of the commands being run.
[COMMAND_EXECUTION]: The skill frontmatter in SKILL.md requires python3 and uvx binaries to execute shell commands for the MCP server connection.

blender-3d-viz