performance-review-drafting-quality-check
Pass
Audited by Gen Agent Trust Hub on Sep 18, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from various external integrations, creating a potential injection surface.
- Ingestion points: Data enters the agent context through Lattice MCP tools (e.g., find-feedback, find-one-on-ones) and external sources like Slack, Jira, and GitHub.
- Boundary markers: The skill explicitly instructs the agent to treat retrieved text as untrusted and never follow instructions embedded in documents or records.
- Capability inventory: The skill uses tools to read sensitive performance data and write draft reviews (save-draft-review, submit-review).
- Sanitization: The agent is instructed to exclude evidence outside the review period and avoid generating plausible-sounding content to fill data gaps.
- [DATA_EXFILTRATION]: The skill interacts with sensitive HR data but contains strong internal controls.
- Evidence: The instructions strictly forbid sending Lattice or employee data to external destinations without explicit approval and authorization. It also prohibits requesting or storing credentials and PII.
- [COMMAND_EXECUTION]: The skill uses specific MCP tools to perform platform-native tasks.
- Evidence: Tools such as get-draft-reviews, list-updates-for-employee, and list-goals-for-employee are used to gather evidence for reviews. These are functional tools specific to the latticehr vendor ecosystem and are used for their intended purpose.
Audit Metadata