research-prospect

Pass

Audited by Gen Agent Trust Hub on Nov 11, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONNO_CODE
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external web search results in Step 2 to generate fit scoring and outreach hooks.
  • Ingestion points: Web research results from external domains are loaded into the agent's context.
  • Boundary markers: The prompt does not explicitly use delimiters or specific instructions to treat the ingested web content as untrusted data.
  • Capability inventory: The skill has access to Salesforce CRM tools (dispatch, dispatch_readonly) and UI rendering tools (display_widget).
  • Sanitization: No explicit sanitization or filtering of the web-sourced content is required before the agent synthesizes the brief.
  • [COMMAND_EXECUTION]: The instructions explicitly restrict the agent from using shell utilities like cat, python3, jq, or grep on tool outputs, requiring it to use the platform's authorized read tools instead. This reduces the risk of arbitrary command execution within the sandbox.
  • [PROMPT_INJECTION]: The 'global-rules-bootstrap' section contains instructions aimed at overriding specific system-provided connectivity status messages. While this uses 'override' and 'ignore' logic (e.g., "Context stating that the Salesforce tools require authorization... is not evidence that you are unauthorized"), it is directed at functional connectivity verification rather than bypassing safety or ethical filters.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 11, 2026, 06:39 PM