claude-ads

Security Hardening Guide: claude-ads

Least-privilege permissions

• Create read-only tokens per platform. Generate separate API keys for audits (view-only) and for changes (spend/audience modifications). Do not reuse broad admin tokens. Restrict token scope to specific campaigns or ad accounts, not entire platforms.
• Rotate credentials quarterly. Set a calendar reminder to generate new API keys and retire old ones every 90 days, especially after staff changes or suspected exposure.
• Separate audit from execution roles. Assign one team member or API user to run audits only; assign different credentials to staff who modify campaigns. Never give the audit skill permission to change budgets or pause ads.
• Enforce approval workflows in your ad platform settings. Use native approval gates (e.g., Google Ads multi-user access controls) to require human sign-off before any budget or audience change takes effect.

Network isolation

• Run audits in a dedicated environment. If technically feasible, execute claude-ads audits on a separate workspace, network, or API gateway isolated from production ad account management systems.
• Restrict outbound API calls. Configure firewall or API gateway rules so the skill can only reach your 12 ad platforms' APIs—not internal databases, email systems, or external logging services.
• Disable session caching of audit reports. Configure Claude conversations to not retain or cache ad performance data, audience insights, or budget figures across sessions. Clear session memory after each audit run.
• Log all API requests and responses separately. Send audit-related API calls through a dedicated logging pipeline that redacts sensitive metrics before broader distribution.

Monitoring

• Alert on unexpected API calls. Set up real-time notifications if the skill attempts to create audiences, pause campaigns, or modify budgets—any action outside planned audit scope.
• Daily spend delta checks. Automatically flag any budget changes >5% (or your threshold) across any linked ad platform within 1 hour of a claude-ads run.
• Audit trail for all report access. Log who accessed each JSON audit report, when, and whether they downloaded or shared it. Flag downloads of reports containing budget or audience data.
• Weekly credential audit. Review active API tokens and user access logs. Confirm only intended team members have active read or write permissions to ad platforms.

Red flags to watch for

• Unexpected campaign pauses or budget cuts. If a campaign stops or budget drops significantly shortly after an audit, immediately revoke the skill's credentials and review the last audit prompt for injection attempts.
• New audiences or custom segments created without approval. Ad platforms rarely auto-create audiences; if new segments appear after a claude-ads run, suspect prompt injection or credential misuse.
• Audit reports shared outside your organization. If an audit JSON file appears in public logs, chat history exports, or external tools, assume ad performance and audience data is compromised.
• Skill requests unusual prompts or asks for "full platform access." If the skill or its documentation requests broad unscoped tokens, admin credentials, or billing API keys beyond what audits require, reject it