opensquilla

Least-privilege permissions

• Restrict data access: Only give OpenSquilla permission to access the specific databases or files it needs for each task—not broad read access to your entire system.
• Limit API credentials: Use separate, low-permission API keys for the agent (e.g., read-only keys) rather than reusing admin credentials.
• Disable external tool calls you don't use: Turn off any integrations or plugins the agent has permission to call if your workflows don't need them.
• Rotate credentials frequently: Change passwords and API keys every 30–90 days, and immediately after staff changes.

Network isolation

• Run in a sandbox or container: Deploy OpenSquilla in an isolated environment (like a Docker container) with limited outbound network access.
• Whitelist external connections only: Block all outbound network traffic by default; only allow connections to approved third-party services your team explicitly needs.
• Air-gap sensitive workflows: For high-risk tasks (e.g., financial or healthcare data), run the agent on an offline network with no internet connection.
• Monitor egress traffic: Log all data leaving your network to detect unusual exports or exfiltration attempts.

Monitoring

• Log all agent decisions and reasoning steps: Keep detailed records of what the agent decided and why—compressed prompts make this harder, so require the agent to "show its work."
• Set up automated alerts: Flag unusual behavior patterns (e.g., repeated permission-denied errors, excessive API calls, or requests for sensitive data outside normal parameters).
• Audit dependency updates: Since OpenSquilla is early-stage, track when its libraries change and review each update for security patches before applying.
• Regular penetration testing: Periodically test the agent with fake injection attacks (malicious prompts disguised as legitimate requests) to confirm defenses work.

Red flags to watch for

• Vague explanations from the agent: If the agent won't or can't explain its reasoning clearly, treat it as a warning—compressed logic may hide unintended behavior.
• Requests for elevated permissions: Watch for patterns where the agent repeatedly asks for higher access levels or tries workarounds when denied.
• Unusual external connections: Any outbound calls to domains or services not on your whitelist should trigger immediate investigation and lockdown.
• Orphaned or unverified dependencies: Before production use, demand that your vendor or team confirms all libraries are actively maintained, have known security histories, and pass dependency scanning tools (like OWASP Dependency-Check).