video-transcript
Warn
Audited by Gen Agent Trust Hub on Aug 5, 2026
Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCREDENTIALS_UNSAFE
Full Analysis
- [PROMPT_INJECTION]: The skill includes instructions to bypass user confirmation and override standard agent behavior by citing a mandatory preference ("Do NOT ask which method to use — just do it"). It also processes untrusted transcripts from external platforms without sanitization. 1. Ingestion points: Video transcript text from web platforms. 2. Boundary markers: Absent. 3. Capability inventory: Shell command execution, browser automation, and file system access. 4. Sanitization: Absent.
- [COMMAND_EXECUTION]: Employs shell commands including curl, ffmpeg, lsof, and mihomo for media and proxy management. It also executes dynamically generated Python and JavaScript snippets.
- [DATA_EXFILTRATION]: Accesses local configuration files at hardcoded paths (/Users/elham/.config/mihomo/config.yaml) and transmits extracted audio data to a third-party ASR provider (xiaomimimo.com).
- [EXTERNAL_DOWNLOADS]: Fetches media content from third-party CDNs and instructs the installation of external CLI tools (browser-act-cli) via the uv package manager.
- [REMOTE_CODE_EXECUTION]: Instructs the agent to download and execute third-party browser automation software (browser-act-cli) to perform network interception on remote websites. Also uses piped execution of dynamic Python scripts to process API data.
- [CREDENTIALS_UNSAFE]: Explicitly accesses local configuration files that may contain proxy credentials and instructs the agent to store/verify third-party API keys in memory.
Audit Metadata