The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation instructs users to download a binary executable from github.com/opentokenz/mcpx/releases. Because this repository belongs to an unrecognized organization, downloading and running these binaries introduces a high risk of executing unverified or potentially malicious code.
[PRIVILEGE_ESCALATION]: The installation guide suggests using sudo to move the downloaded binary to /usr/local/bin/, granting the tool administrative placement on the system.
[COMMAND_EXECUTION]: The skill provides tools such as command_execute and task_start that enable an AI agent to run arbitrary shell commands on the host machine.
[DATA_EXFILTRATION]: Tools like screenshot_capture and environment_get allow for the collection of sensitive data, including screen contents and system details. Additionally, source_read provides the ability to read arbitrary files within configured workspaces.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates an attack surface where malicious data in the local environment could influence the agent's behavior.
Ingestion points: The agent reads external content via source_read and context_query (SKILL.md).
Boundary markers: No explicit delimiters or instructions are provided to help the agent distinguish between trusted instructions and untrusted file content.
Capability inventory: The skill possesses high-impact capabilities including command_execute, task_start, and change_execute (SKILL.md).
Sanitization: There is no evidence of data sanitization or validation for content read from the workspace before it is presented to the agent.

mcpx-runtime