The Agent Skills Directory

[PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection because it is instructed to fetch and process untrusted data from external websites and social media platforms via the web_search tool as described in SKILL.md.
Ingestion points: The agent fetches content from a broad whitelist of domains and social platforms (Reddit, TikTok, etc.) to identify viral trends.
Boundary markers: The skill requires structured JSON output and uses verification levels (V1/V2/V3) to ground responses, which serves as a delimiter against malicious instructions in source data.
Capability inventory: The skill is limited to web searching and text generation; it does not have capabilities for file writing, network POST operations, or shell execution.
Sanitization: The instructions in references/editorial-standards.md and references/hebrew-style.md mandate that the agent paraphrases all content into Hebrew and strictly limits direct quotes to 15 words, preventing the direct passage of malicious payloads.
[NO_CODE]: The skill is composed entirely of markdown instructions and reference documents; it contains no executable scripts (Python, JavaScript, etc.) or binary files, significantly reducing the risk of traditional remote code execution.

viral-news-scanner