Data Exfiltration Review

Data Exfiltration Review

Core Workflow

1. Identify what data the skill can access directly or through scripts, MCP tools, host permissions, examples, or references.
2. Identify sinks: network calls, external APIs, logs, generated files, shell output, issue comments, customer messages, and copied reports.
3. Check for hidden or unnecessary data movement, broad globbing, credential reads, encoded payloads, or unexplained uploads.
4. Verify the workflow uses least privilege and only accesses data needed for the user-requested artifact.
5. Recommend removals, redaction, allowlists, read-only defaults, and approval gates.
6. Report residual risk and safe-use conditions.

Safety Rules

• Do not inspect or print secrets unless the user explicitly asks for a secrets audit and a safe handling path exists.
• Do not approve outbound network behavior without a clear purpose.
• Do not allow skills to collect broad local data unrelated to the task.

Deliverable Shape

For exfiltration reviews, provide:

• Data sources
• Data sinks
• Unnecessary or hidden data movement
• Secrets and sensitive data concerns
• Least-privilege recommendations
• Approval conditions
• Pass/warn/fail/block recommendation

References

• Read references/data-exfiltration-review-checklist.md when reviewing data access, output, logs, scripts, or network behavior.