skills/sshvarzman-salesforce/claude-skills/authoring-record-summary-prompt-templates/Gen Agent Trust Hub
authoring-record-summary-prompt-templates
Pass
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data originating from Salesforce records and conversation transcripts, creating a surface for indirect prompt injection.
- Ingestion points: The PromptFlow in
assets/inh-record-summary-context.flow-meta.xmlingests data fromCase.Description,Contact.Description,Account.Description, and external conversation transcripts via thegetConvTscpForRecordaction. - Boundary markers: The instructions in
SKILL.mdand the metadata inassets/inh-record-summary-template.genAiPromptTemplate-meta.xmlinclude explicit grounding instructions such as "Use ONLY the facts below; never invent" and "Using ONLY the context above". These act as negative constraints to help prevent the model from following instructions embedded in the data. - Capability inventory: The skill is limited to data retrieval (
recordLookups) and prompt assembly. It lacks capabilities for unauthorized file writes, privilege escalation, or external network exfiltration. - Sanitization: The skill relies on natural language boundary instructions within the prompt to sanitize the input. While these instructions direct the LLM to ignore outside information, they do not provide technical sanitization against sophisticated adversarial inputs.
Audit Metadata