The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The helper script 'generate_video_transcript.sh' accesses local credential stores to facilitate authenticated downloads. It attempts to read session cookies from the Linear desktop application's SQLite database located at '/Library/Application Support/Linear/Cookies'. It also retrieves GitHub authentication tokens using the 'gh auth token' command or by reading 'GITHUB_TOKEN' and 'GH_TOKEN' environment variables.\n- [COMMAND_EXECUTION]: The skill executes local commands to process data and configuration. It runs a Python script to parse the Linear cookie database and sources the user's '/.bash_profile' to load API keys and environment settings.\n- [DATA_EXFILTRATION]: The skill accesses sensitive information from the local file system, including shell profiles and application cookies, and performs network operations. While these operations are directed at well-known services (GitHub, Linear, and Google Gemini API) to fulfill the skill's stated purpose, the pattern of accessing local credentials and sending data externally represents a significant risk profile.\n- [PROMPT_INJECTION]: The skill features an indirect prompt injection surface where the user-provided video title is interpolated into the prompt sent to the LLM. While the title is escaped for XML attribute syntax, the absence of clear boundary markers or instructions to ignore embedded commands creates a risk that malicious input could influence the model's behavior.\n- [EXTERNAL_DOWNLOADS]: The script downloads video files from external sources such as 'uploads.linear.app' and GitHub's attachment infrastructure to process them for transcription.

video-transcripts