The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the exec tool to perform administrative tasks for the gbrain utility. All critical commands, such as gbrain self-upgrade and gbrain config set, are hardcoded within the skill instructions. This design pattern ensures that the agent does not execute arbitrary or malicious commands received from external sources.
[INDIRECT_PROMPT_INJECTION]: The skill identifies a potential attack surface where untrusted data (version markers, changelogs, and migration playbooks) could contain malicious instructions.
Ingestion points: stderr output from the gbrain command, JSON responses containing changelogs, and remote markdown files from github.com/garrytan/gbrain.
Boundary markers: The skill instructs the agent to summarize external data into bullets rather than reflecting raw content, providing a level of sanitization.
Capability inventory: Access to the exec tool for shell command execution.
Sanitization: The instructions explicitly forbid executing any command parsed from external text, mandating that the agent only use hardcoded command strings.
[EXTERNAL_DOWNLOADS]: The skill fetches configuration and migration playbooks from the vendor's official GitHub repository (github.com/garrytan/gbrain). These downloads are used to inform the upgrade process and do not involve the direct execution of remote scripts.
[DATA_EXFILTRATION]: No evidence of sensitive data access or unauthorized exfiltration was found. The skill's network activity is limited to legitimate update checks and retrieving vendor-provided documentation.

gbrain-upgrade