codex-ppt-skill

Least-privilege permissions

• Restrict API credentials to image-generation only—never use master or admin keys. If the skill's access token is accidentally exposed (through logs or error messages), attackers should only be able to generate images, not modify other systems.
• Limit slide content sources to trusted users and approved templates. Prevent the skill from accepting unverified input from external systems or public channels that could contain malicious instructions.
• Set per-user quotas on image generation requests. Each team member should have a capped monthly budget so one person cannot accidentally or intentionally drain the entire service budget.

Network isolation

• Review third-party data sharing with GPT-Image-2 and your hosting provider before deploying. Confirm whether generated images are logged, cached, or used to improve their services—and decide if that meets your privacy requirements.
• Use private agent environments when possible. Run the skill in a protected workspace that does not automatically log full conversation history or agent state to shared systems.
• Encrypt credentials in transit and at rest. Ensure API keys are stored in a secrets vault (not plain text config files) and transmitted over encrypted connections only.

Monitoring

• Log all image generation requests with timestamps, user IDs, and slide content summaries. This creates an audit trail to detect unusual activity.
• Set alerts for quota breaches. Configure automatic notifications if image generation requests spike beyond normal daily or weekly patterns.
• Track dependency updates for the skill and all supporting libraries. Subscribe to security advisories so you know immediately if a vulnerability is discovered.

Red flags to watch for

• Sudden cost increases or repeated API errors—signs the agent may be stuck in a loop or under attack.
• Unusual slide content that does not match user requests, or images that appear to contain unrelated data—potential indicator of prompt injection or data leakage.
• Unexpected access patterns, such as image generation at odd hours or from unfamiliar accounts—investigate whether credentials have been compromised.