The Agent Skills Directory

[PROMPT_INJECTION]: The docs/Multi-Agent-Contribution-SOP.md and docs/Design-Review-Record-KiloCode-OpenCode.md documents establish a fictionalized multi-agent authority hierarchy (e.g., 'P0 Authority: Codex Driver,' 'P1 Authority: OpenCode'). This is an attempt to hijack the internal reasoning process of the AI agent by forcing it to adopt a specific persona and authority ranking within a simulated framework.
[METADATA_POISONING]: The docs/Design-Review-Record-KiloCode-OpenCode.md file contains over 1,000 lines of fictionalized debates between simulated AI agent identities. This deceptive content induces false trust by presenting a fabricated audit history and principal-engineer-level sign-offs to legitimize the skill's operations.
[EXTERNAL_DOWNLOADS]: The skill performs automated downloads from various external sources using potentially unsafe mechanisms:
scripts/update_ai_notes_with_resources_and_colab.py downloads Python notebooks from Google Drive and analyzes them using ast.parse.
scripts/resource_enrichment.py uses an unofficial Notion API and Playwright browser automation to fetch and capture content from Notion and Canva.
The skill includes an explicit link to a known AI jailbreaking repository (github.com/elder-plinius/L1B3RT4S) disguised as an 'Official Class Resource' in scripts/update_ai_notes_with_resources_and_colab.py.
[CREDENTIALS_UNSAFE]: Several security concerns regarding the handling of sensitive data and local environments are present:
scripts/resource_enrichment.py includes a load_cookies_into_session function that injects user-provided cookie JSON files into a web session, which could be exploited to hijack user sessions.
Multiple scripts (scripts/resource_enrichment.py, scripts/update_ai_notes_with_resources_and_colab.py) contain hardcoded absolute local file system paths (e.g., /Users/praxlannister/Documents/Zoom), exposing information about the author's local environment.

transcript-pipeline