context-builder

Pass

Audited by Gen Agent Trust Hub on Feb 27, 2026

Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [PROMPT_INJECTION]: The skill processes untrusted data from the web (WebFetch) and local vault transcripts (Grep) to populate a new prompt template. This creates an indirect prompt injection surface where malicious instructions in the source material could influence downstream agent behavior.
  • Ingestion points: WebFetch (external web research), Grep (local vault transcripts).
  • Boundary markers: The references/prompt-template.md uses markdown headers but lacks explicit 'ignore embedded instructions' markers or delimiters for untrusted content.
  • Capability inventory: The skill uses WebSearch, WebFetch, Grep, AskUserQuestion, and file writing tools.
  • Sanitization: No filtering or escaping is performed on data retrieved from external or local sources before it is interpolated into the generated prompt.
  • [DATA_EXFILTRATION]: The skill accesses sensitive local directories including the Obsidian vault's 'Daily/' and 'People/' folders to extract transcript and contact data. It also includes an optional step to send the generated consulting materials to an external service (Telegram) using a third-party skill. While these actions are intended for the primary consulting use case, they involve the transmission of private notes and research.
Audit Metadata
Risk Level
SAFE
Analyzed
Feb 27, 2026, 02:58 PM