The Agent Skills Directory

[DATA_EXFILTRATION]: The instructions in references/design-picker.md direct the agent to launch a web server using python3 -m http.server 8723. By default, this server binds to all available network interfaces (0.0.0.0), which can expose the entire project directory to other devices on the same network. This presents a high risk if the directory contains sensitive information such as environment variables (.env) or source control metadata (.git).\n- [EXTERNAL_DOWNLOADS]: The script scripts/package-loader.mjs includes functionality to dynamically download and install NPM packages (specifically @hyperframes/producer and sharp) during execution. Although the script attempts to mitigate risks by requiring pinned versions and disabling lifecycle scripts via --ignore-scripts, the dynamic fetching and installation of external code at runtime remains a potential attack vector.\n- [COMMAND_EXECUTION]: The skill relies on executing various system commands and CLI tools, including python3, node, npx, curl, and kill. While these are used for standard development workflows like previewing and rendering, the broad use of shell commands increases the potential surface area for command injection if user-supplied data is not correctly handled.\n- [INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves processing external transcription and audio data to generate video content. This creates a surface for indirect prompt injection, where malicious instructions hidden in transcripts or metadata could potentially influence the agent's behavior during the planning or composition phases.\n
Ingestion points: Transcription files (transcript.json, .srt, .vtt) and audio data files (audio-data.json) as described in references/transcript-guide.md and references/audio-reactive.md.\n
Boundary markers: The skill instructions provide some guidance for cleaning data but lack explicit 'ignore' instructions or strict boundary markers for the data it processes.\n
Capability inventory: The skill utilizes Puppeteer for headless rendering, performs file system writes, and executes shell commands across multiple scripts.\n
Sanitization: Basic word filtering is suggested in references/transcript-guide.md, but there is no comprehensive sanitization to prevent injection attacks via processed audio or text data.

hyperframes