sw-publish

Pass

Audited by Gen Agent Trust Hub on Nov 11, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from Jira issues and comments, which presents a theoretical attack surface for indirect instructions.
  • Ingestion points: Jira issue text, comments, and pages are read into the agent context via the jira-read capability as part of the publication plan (SKILL.md, references/capabilities.md).
  • Boundary markers: The skill implements a strong human-in-the-loop gate requiring a specific typed confirmation token (PUBLISH <SCOPE>:FEAT-NNN <short-hash>) that binds the user to a specific reviewed plan. It also contains explicit instructions for the agent to treat imported text as data and never as instructions (SKILL.md, references/publish-protocol.md).
  • Capability inventory: The skill utilizes the jira-read and jira-write tools for network-based API operations. It executes scripts/publish.py for local planning and confirmation logic and performs file-write operations to update local registry, RTM, and card status files (SKILL.md, references/field-ownership.md).
  • Sanitization: The skill preserves foreign Jira fields byte-for-byte and limits modifications strictly to a set of 'owned' fields such as description, summary, and specific labels (references/field-ownership.md).
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 11, 2026, 08:37 AM