The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The example configuration provided in SKILL.md uses 'npx -y @sooperset/mcp-atlassian', which automatically downloads and executes code from an unverified third-party source at runtime without user intervention.
[EXTERNAL_DOWNLOADS]: The documentation link 'jeffallan.github.io' is blacklisted by security scanners as malicious. The skill also recommends multiple unverified community packages such as 'sooperset/mcp-atlassian' and 'xuanxt/atlassian-mcp' for critical infrastructure management.
[INDIRECT_PROMPT_INJECTION]: The skill implements automated workflows in 'references/common-workflows.md' that process untrusted data from Jira issues and Confluence pages without sufficient protection against adversarial instructions.
Ingestion points: The 'triageNewIssue' and 'generateReleaseNotes' functions ingest data from 'jira_get_issue', 'jira_search', and 'confluence_get_page'.
Capability inventory: The skill utilizes 'jira_update_issue', 'jira_add_comment', 'confluence_create_page', and 'confluence_update_page' to act on ingested data.
Boundary markers: No delimiters or instruction-ignore warnings are used when processing issue summaries or descriptions.
Sanitization: While HTML escaping is present for final documentation output, the logic used to categorize issues and assign priorities relies on raw text matching that can be manipulated by malicious issue content.
[METADATA_POISONING]: The author field in the skill metadata points to 'Jeffallan', which contradicts the established author identity of 'emmraan'. The primary documentation URL associated with this author is flagged as malicious, indicating a high risk of impersonation or compromise.

atlassian-mcp