The Agent Skills Directory

[DATA_EXFILTRATION]: The skill provides instructions for reading sensitive local files and transmitting their contents over the network. Specifically, it includes examples for reading private SSH keys (e.g., using @id_rsa) to set project variables or add deploy keys. This sequence matches a data exfiltration pattern where sensitive local data is sent to a remote GitLab instance.\n- [CREDENTIALS_UNSAFE]: The skill documents commands for generating and managing persistent authentication tokens and sensitive project variables. It includes instructions for creating personal, project, and group access tokens (glab token create). While it contains warnings against logging these tokens, the ability to generate new long-lived credentials via an agent increases the risk of unauthorized persistence and credential manipulation.\n- [COMMAND_EXECUTION]: The skill enables a broad range of system operations via the glab CLI, including the ability to make arbitrary REST and GraphQL requests via glab api. This provides an extensive interface for interacting with remote systems and internal project configurations.\n- [PROMPT_INJECTION]: The skill contains an indirect prompt injection attack surface because it retrieves and processes content from external sources.\n
Ingestion points: The agent is instructed to fetch data from GitLab issues, merge requests, CI/CD logs, and repository files via commands like glab mr view, glab issue view, and glab ci trace.\n
Boundary markers: No instructions are provided to wrap this external content in delimiters or to advise the agent to ignore instructions embedded within the retrieved data.\n
Capability inventory: The skill possesses the ability to perform network operations, write to the local file system (e.g., glab release download), and manage credentials (glab token), which could be abused if the agent follows malicious instructions from ingested data.\n
Sanitization: No validation or sanitization steps are documented for the content retrieved from GitLab before it is processed by the agent.

glab