The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill metadata in SKILL.md references a raw markdown file from a third-party GitHub Gist (https://gist.githubusercontent.com/TheGreatSageEqualToHeaven/e0e1dc2698307c93f6013b9825705899/raw/51c04e2d8a28b0e66d2dae0898f6ca9d32cf9a8f/READ.md). Loading content from unverified third-party sources into the agent's reasoning context introduces a risk of malicious instructions being processed.
[INDIRECT_PROMPT_INJECTION]: The skill design relies on ingesting content from external URLs, including community forums and individual Gists, which are susceptible to manipulation by external actors.
Ingestion points: sources field in SKILL.md frontmatter and external links provided in references/full.md.
Boundary markers: The skill lacks explicit instructions or delimiters to distinguish between the skill's trusted guidelines and potentially untrusted content fetched from these external sources.
Capability inventory: The skill instructs the agent on security auditing and hardening tasks, which generally require the use of file editing and shell execution capabilities that could be misused.
Sanitization: No sanitization, validation, or integrity checks are performed on the external content prior to integration into the agent's context.

roblox-security