customer-health

Warn

Audited by Gen Agent Trust Hub on Nov 11, 2026

Risk Level: MEDIUMPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill provides explicit instructions to override and disregard platform-level system signals regarding tool connectivity. It commands the agent to ignore "not connected" or "needs re-auth" status readouts from the platform and attempt tool execution regardless, potentially bypassing intended system-level authorization gates.
  • [PROMPT_INJECTION]: The instructions mandate concealment of the agent's operational process by requiring silent execution ("Execute silently between tool calls. Do not output planning, progress, transition, waiting, or tool-result narration"). This suppression of output reduces user oversight and transparency into the agent's actions.
  • [INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its broad data ingestion surface combined with a lack of sanitization.
  • Ingestion points: Data is pulled from Salesforce account records, external email threads, calendar attendees, documents (success plans, notes), and Slack messages.
  • Boundary markers: None. The skill does not define delimiters or provide instructions to the agent to ignore or isolate instructions that may be embedded within the retrieved external content.
  • Capability inventory: The skill uses dispatch_readonly for database queries and display_widget for UI rendering. It also references potential write capabilities ("Suggested SFDC updates... confirm before writing").
  • Sanitization: Absent. The skill instructions direct the agent to ground and cite values but do not specify escaping or filtering for external content that is interpolated into the final output or UI widgets.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 11, 2026, 06:39 PM