The Agent Skills Directory

Indirect Prompt Injection (MEDIUM): The skill is designed to build systems that ingest untrusted data from external chat platforms.\n
Ingestion points: Untrusted data enters via message.content in references/moderation-system.md (AutoMod) and via Update.de_json(data) in references/platform-templates.md (Telegram webhook).\n
Boundary markers: Code snippets lack explicit boundary markers or delimiters when interpolating user-provided chat strings into logic flows.\n
Capability inventory: The skill requests high-privilege tools including Bash, Write, Edit, and WebFetch to manage bot infrastructure.\n
Sanitization: While the documentation provides a security checklist recommending input sanitization, the provided implementation snippets (e.g., the fuzzy-matching word filter) are susceptible to adversarial inputs designed to bypass filters and potentially influence an agent monitoring the bot's logs.\n- Command Execution (LOW): The skill requests the Bash tool. While no malicious shell scripts are found, the combination of bash access and external data ingestion increases the overall attack surface.\n- Credentials Unsafe (INFO): Hardcoded placeholder credentials (postgresql://user:pass@localhost/botdb) are present in references/platform-templates.md for demonstration purposes.

bot-developer