The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from public sources such as HackerNews, arXiv, TLDR AI, The Batch, and GeekNews, incorporating this content directly into synthesis prompts. An attacker could craft malicious content in these public feeds to influence the agent's output or downstream actions.
Ingestion points: Data is collected from various third-party APIs, RSS feeds, and HTML pages as detailed in SKILL.md and SKILL_detail.md.
Boundary markers: The synthesis prompt uses an [Insert collected data] placeholder and relies on natural language instructions to separate data from instructions. It lacks robust structural delimiters or technical sanitization.
Capability inventory: The skill possesses the Bash tool (for network and text processing), Write (for saving local files), and WebSearch capabilities.
Sanitization: The skill includes an explicit warning to the agent to treat collected content as data, not instructions, and enforces a human-in-the-loop gate before performing automated chaining actions.
[COMMAND_EXECUTION]: The skill uses the Bash tool to perform complex text extraction using grep and sed on data fetched from the web. It also probes for the presence of external binaries such as yt-dlp, ffmpeg, and curl_cffi for video processing tasks.
[EXTERNAL_DOWNLOADS]: The skill performs frequent network requests to external domains including hn.algolia.com, arxiv.org, tldr.tech, deeplearning.ai, and news.hada.io to gather trends. These are legitimate sources for the skill's stated purpose.

frontier-digest