buaa-classroom-summarizer

Warn

Audited by Gen Agent Trust Hub on Sep 21, 2026

Risk Level: MEDIUMDATA_EXFILTRATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill accesses local Chromium-based browser cookie databases (Microsoft Edge and Google Chrome) to extract authentication tokens for the domain buaa.edu.cn. It utilizes the Windows CryptUnprotectData API to decrypt these cookies for use in authenticated network requests.
  • Evidence: scripts/extract_buaa_classroom.py accesses paths like AppData/Local/Microsoft/Edge/User Data/Default/Network/Cookies and uses win32crypt.CryptUnprotectData to retrieve session values.
  • [DATA_EXFILTRATION]: Extracted session tokens and authorization headers are stored in a local cache file at ~/.codex/cache/buaa_browser_session.json in plain text, potentially exposing the user's active session to other local processes.
  • [COMMAND_EXECUTION]: The skill uses subprocess.run to execute internal scripts and external binaries like ffmpeg for video processing and OCR extraction.
  • Evidence: scripts/collect_buaa_course_replays.py and scripts/extract_buaa_classroom.py invoke subprocess.run to handle video and PPT outline extraction tasks.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, specifically automated transcripts and OCR content from classroom videos, which are then passed to the LLM for summarization.
  • Ingestion points: metadata.json, transcript.txt, and ppt_outline.md are generated from BUAA API responses and video OCR.
  • Boundary markers: The skill uses specific prompts (semantic_rebuild_prompt.md) and instructions to guide the agent, though it lacks robust structural delimiters for the untrusted content.
  • Capability inventory: The agent has the ability to read/write files and execute command-line tools within the platform environment.
  • Sanitization: A validation script (scripts/validate_final_note.py) is used to check for the presence of raw artifacts or boilerplate in the final output before it is marked as complete.
  • [DYNAMIC_EXECUTION]: The skill dynamically loads Python modules from local file paths using the importlib library.
  • Evidence: scripts/collect_buaa_course_replays.py loads the sister script extract_buaa_classroom.py using importlib.util.spec_from_file_location.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 21, 2026, 04:01 AM