adr
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The ADR review workflow described in 'SKILL.md' (Section 5) facilitates an indirect prompt injection attack surface by processing untrusted external data.
- Ingestion points: The skill fetches ADR content from Confluence (an external and potentially attacker-controlled source) using
scripts/confluence-get.ts. - Boundary markers: The prompt provided to the 'ultrabrain' sub-agent interpolates the fetched content directly into the prompt instructions without using delimiters or instructions to ignore embedded commands.
- Capability inventory: The skill possesses capabilities to read, create, and update pages in the Confluence 'CE' space and execute local shell commands via
bunx. - Sanitization: The instructions specify converting HTML to readable text, but do not describe any sanitization, filtering, or escaping of the content before it is processed by the AI sub-agent.
- [COMMAND_EXECUTION]: The instructions in 'SKILL.md' rely on executing local TypeScript scripts using
bunx tsxandnpx tsx(e.g.,scripts/adr-report.ts,scripts/adr-sync-emojis.ts). While these scripts are part of the skill's package and perform intended administrative tasks, the reliance on shell execution for routine operations increases the local attack surface.
Audit Metadata