cc-sdd

Security Hardening Guide: cc-sdd Agent Skill

Least-privilege permissions

• Grant only required access: Configure the agent to access only the specific files, systems, and tools needed for the current task—not blanket admin rights or broad folder access.
• Use temporary credentials: Issue time-limited login credentials that expire after the job completes, rather than permanent keys that could be reused if stolen.
• Separate spec approval from execution: The person who approves the specification should not be the same account that runs the implementation; this prevents a single compromised credential from both authorizing bad work and executing it.
• Audit permission changes: Log any request by the agent to expand its access level; require manual human approval before granting additional permissions.

Network isolation

• Restrict outbound connections: Block the agent from sending data to external cloud services, email, or the internet unless explicitly necessary for the task (e.g., pulling approved libraries).
• Use air-gapped or VPN environments: Run long-running agents on isolated networks or behind a VPN so they cannot phone home or reach unexpected destinations.
• Whitelist dependencies only: If the agent needs to fetch code libraries, pre-approve a whitelist of safe sources; do not allow it to download from any repository it finds.

Monitoring

• Log all code generation and deployment: Record what code the agent created, what system changes it made, and when—store logs where they cannot be deleted by the agent itself.
• Real-time execution alerts: Set up automated notifications if the agent attempts unusual actions (e.g., accessing files outside the approved project, making network calls, requesting new credentials).
• Regular credential and log audits: Weekly or monthly review of who accessed what and whether any permissions were misused; look for patterns like repeated failed attempts to escalate access.

Red flags to watch for

• Specification documents that contain code or commands: Reject specs that include hidden instructions or executable content; a spec should describe what to build, not how to attack the system.
• Agent requests for broader permissions mid-task: If the agent asks for access to databases, user files, or external services after approval, pause and investigate why before granting.
• Unexplained network traffic or file transfers: Stop execution if logs show the agent uploading code, configuration files, or logs to unknown external locations.
• Mismatches between approved spec and generated code: If the final code does significantly more, accesses different systems, or includes external dependencies not mentioned in the spec, require human security review before deployment.