The Agent Skills Directory

[METADATA_POISONING]: The skill-report.json file contains a pre-calculated security audit section that explicitly labels the skill with a 'safe' risk level and attempts to dismiss potential security findings as false positives. This self-referential content is a deceptive practice intended to influence the safety evaluation of the skill.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its broad data ingestion capabilities combined with active file system tools.
Ingestion points: The skill fetches content from external links, shared documents, and messaging platforms such as Slack, Microsoft Teams, Google Drive, and SharePoint through platform integrations and MCP servers (SKILL.md).
Boundary markers: None. The instructions lack explicit delimiters or 'ignore embedded instructions' directives when processing external context.
Capability inventory: The skill uses create_file to generate document scaffolds, str_replace for iterative drafting and surgical edits, and invokes sub-agents for automated 'Reader Testing' (SKILL.md).
Sanitization: None identified. Data retrieved from untrusted external sources is processed and interpolated into the workspace without validation or escaping.

doc-coauthoring