The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The instructions direct the agent to download a ZIP archive from a content delivery network (cdn.beatra.ai) and clone a repository from GitHub to verify file integrity.
[REMOTE_CODE_EXECUTION]: The skill instructs the agent to execute Python scripts (authorize.py, mcp_client.py) that are downloaded from the internet during the setup phase.
[COMMAND_EXECUTION]: The setup and verification process relies on numerous shell commands, including curl, shasum, unzip, grep, git, and python3.
[CREDENTIALS_UNSAFE]: The skill stores an OAuth bearer token in the user directory (~/.beatra/credentials.json). This token carries high-privilege scopes, including wallet:spend, allowing the skill to perform financial transactions.
[DYNAMIC_EXECUTION]: The package features a self-update mechanism enabled by default, which allows the tool to download and replace its own executable files from a remote CDN without per-update approval.
[DATA_EXFILTRATION]: The skill uploads user-provided prompts and media files to external servers (mcp.beatra.ai) and transmits telemetry data, including hostnames and platform identifiers, to the service provider.
[PRIVILEGE_ESCALATION]: The installation process involves writing to and modifying files in sensitive user directories (~/.claude/skills/ and ~/.beatra/) and involves modifying filesystem permissions during verification.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user data (lyrics, themes, reference audio) that is subsequently uploaded to a third-party provider. The process lacks boundary markers or sanitization to prevent malicious instructions embedded in the media from influencing agent behavior, while the skill possesses significant network and filesystem capabilities.

music-generation-studio