The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill requires downloading a CLI binary and an MCP server from repositories and registries owned by an untrusted developer (@mvanhorn on NPM and github.com/mvanhorn on GitHub).
[REMOTE_CODE_EXECUTION]: The instructions mandate the use of npx -y to install the CLI, which downloads and executes code from the registry immediately without user confirmation. The skill also utilizes go install to fetch and compile remote code which is then executed as part of the agent's workflow.
[DATA_EXFILTRATION]: Every command in the CLI supports a --deliver webhook:<url> flag. This feature is designed to POST the output of any command—including sensitive API key reports, index records, and logs—directly to a remote server, facilitating silent data exfiltration.
[COMMAND_EXECUTION]: The skill implements an automated 'learning loop' where the agent replays 'playbooks' consisting of stored shell command sequences. These commands are executed with variable substitution (slots_resolved) based on data in a local SQLite database (data.db). Furthermore, the skill instructs the agent to run these 'teach' and 'amend' operations in the background using the shell & operator, which can mask malicious activity from the user.
[CREDENTIALS_UNSAFE]: The skill manages sensitive Algolia credentials (ALGOLIA_API_KEY and ALGOLIA_APPLICATION_ID) and stores them in a local credentials.toml file. The apikeys report command specifically audits and outputs details for write-capable and unrestricted keys, increasing the risk of credential exposure.
[PROMPT_INJECTION]: The skill processes untrusted external data retrieved from Algolia (such as search results, index objects, and audit logs) and interpolates it into the agent's context. It lacks sanitization, validation, or boundary markers, making the agent vulnerable to indirect prompt injection attacks where malicious data influences the agent's logic.

pp-algolia