The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple sources including GitHub issue titles, descriptions, and PR comments via the GitHub MCP tools. It also supports reading image files from disk for accessibility analysis. Maliciously crafted content in these external sources could influence the agent's behavior or be included in the generated reports.
Ingestion points: Retrieving data using mcp_github_github_search_issues and mcp_github_github_issue_read, and reading image files via chat.imageSupport.enabled (SKILL.md).
Boundary markers: The skill instructions lack specific delimiters or warnings to the agent to ignore instructions embedded within the GitHub issue content or image analysis results.
Capability inventory: The skill writes report files to the local workspace at .github/reviews/accessibility/ and generates HTML output. It also has the ability to search across all repositories and organizations the user has access to (SKILL.md).
Sanitization: There is no evidence of sanitization, filtering, or validation of the external GitHub content or image metadata before it is interpolated into workspace documents or the conversation.

insiders-a11y-tracker