The Agent Skills Directory

[DATA_EXFILTRATION]: The skill extracts audio from user videos and uploads it to https://uguu.se/upload. This is a public, anonymous file-hosting service. Uploading potentially sensitive user data to an untrusted third-party platform represents a high risk of data exposure.
[COMMAND_EXECUTION]: The server scripts review_server.js and subtitle_server.js utilize child_process.execSync to run shell commands like ffmpeg and ffprobe. These commands are constructed using file paths and arguments passed directly from the environment or user input without sufficient sanitization, creating a significant attack surface for command injection.
[EXTERNAL_DOWNLOADS]: The workflow relies on the Volcengine ASR API for transcription. While this is a known commercial service, the skill also utilizes anonymous hosting services to bridge data for the API, involving the transfer of local data to remote servers outside the user's control.
[PROMPT_INJECTION]: The skills/自进化/ module implements a 'self-evolution' mechanism where the agent is instructed to update its own logic and rule files based on context and user feedback. This creates a vulnerability to indirect prompt injection; malicious data processed by the agent (such as a poisoned video transcript) could result in the permanent alteration of the skill's instructions or security configuration.
Ingestion points: Transcription results from Volcengine API in volcengine_result.json.
Boundary markers: None identified in processing scripts.
Capability inventory: Shell command execution via ffmpeg and file system writes in review_server.js.
Sanitization: No escaping or validation is performed on the transcription data before processing.

videocut