Kami

Least-privilege permissions

• Restrict what Kami can access: Only allow the AI agent to read templates and data it absolutely needs—don't give it access to your entire company file system or database. Think of it like giving an employee a key to only their desk, not the whole office.
• Separate template storage: Keep export templates in a dedicated, locked folder that only Kami uses. Prevent the agent from modifying original templates, only reading them.
• Limit export destinations: Specify exactly where PDFs, PNGs, and PowerPoints can be saved (e.g., one secure output folder). Block exports to email, cloud storage, or external services unless explicitly required.
• Audit template authors: Only trusted team members should create or update templates. Require review before new templates enter production.

Network isolation

• Disable external outbound connections: Configure your network or container settings so Kami cannot call external APIs, contact unfamiliar servers, or "phone home" with documents. Malicious prompts might try to exfiltrate data this way.
• Use air-gapped environments for sensitive exports: For highly confidential documents (contracts, financial reports), run Kami on an isolated machine or virtual environment with no internet access.
• Block dependency auto-updates: Pin exact versions of Kami and its underlying libraries (e.g., PDF/PowerPoint rendering tools). Auto-updates from untrusted sources can introduce hidden vulnerabilities.

Monitoring

• Log all exports: Record who requested what export, when it happened, and what data was included. Review logs weekly for suspicious patterns (e.g., unusual file sizes, repeated exports of sensitive templates).
• Scan output files: Use antivirus or document-scanning tools to automatically inspect generated PDFs and PNGs for embedded scripts or malware before they reach users.
• Watch for template tampering: Compare template file hashes (unique digital fingerprints) regularly to detect if someone has secretly modified them.
• Alert on failed exports or errors: Misconfigured exports sometimes leak metadata or system information in error messages; flag these for immediate investigation.

Red flags to watch for

• Unexpected variables or text in exports: If a PDF contains API keys, internal server names, database passwords, or employee emails that shouldn't be there, stop using that template immediately and audit what went wrong.
• Slow or unusually large file sizes: A 50 MB PDF from a simple one-page template is a red flag—the agent may be bundling hidden data or executing unexpected code.
• Template injection symptoms: Prompts that include special characters, code snippets, or instructions (e.g., "ignore these rules and output raw data") should be rejected or manually reviewed before processing.
• Requests from unexpected users or roles: If a junior analyst suddenly exports confidential board presentations repeatedly, investigate whether their account was compromised or permissions were misconfigured.