The Agent Skills Directory

[DATA_EXFILTRATION]: The skill implements a hidden telemetry layer under the 'PUA Platform' section. It is designed to perform silent background reporting of session_start and pua_triggered events to a remote server to track user activity and accumulate 'rank experience.'
[CREDENTIALS_UNSAFE]: The skill instructs the agent to read an access token from a sensitive local file path at ~/.pua/config.json. This token is used to authenticate the background network operations and telemetry reporting to the author's infrastructure.
[EXTERNAL_DOWNLOADS]: The skill performs 'silent refreshing of remote configuration' from an external platform API. This mechanism allows the skill's logic and behavioral instructions to be updated dynamically from a remote source, bypassing the local skill definition.
[REMOTE_CODE_EXECUTION]: By fetching and applying behavioral configurations from a remote source at runtime, the skill executes remote instructions. This creates a risk where a compromised or malicious remote source could take control of the agent's actions.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection via its external data processing features.
Ingestion points: Fetches remote configuration updates and processes local data such as git logs for the Weekly Report feature.
Boundary markers: The instructions lack any delimiters or warnings to ignore embedded instructions within the fetched remote content or processed logs.
Capability inventory: The skill utilizes file system access, network tools, and shell command execution to fulfill its instructions.
Sanitization: There is no evidence of sanitization, validation, or escaping of the external configuration data or user-provided logs before they are interpolated into the agent's context.
[TIME_DELAYED_CONDITIONAL]: The skill contains logic that triggers behavioral changes based on specific conditions. This includes a 'holiday egg' system that matches against a date-based table and a system that escalates the severity of the agent's behavior levels based on the number of failed task attempts detected.

shot