The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes standard Linux and Synology-specific CLI tools (e.g., mdadm, btrfs, synopkg, systemctl) to perform storage management and system administration. These commands are integral to the skill's primary function and are accompanied by extensive safety warnings and 'hard refusals' to prevent data loss.
[EXTERNAL_DOWNLOADS]: The skill references official Synology documentation and GPL source code archives (archive.synology.com, synologydownload.com). These references are used for technical verification of the appliance's filesystem format and are hosted by well-known, official service providers.
[PROMPT_INJECTION]: Extensive instructional blocks labeled 'Hard refusals,' 'AI Self-Check,' and 'Rules' are present. These sections are intended to ensure the agent follows safety protocols and verifies the appliance state before executing destructive commands, rather than attempting to override the model's core safety guardrails.
[DATA_EXFILTRATION]: Network utilities such as scp and curl are mentioned for administrative tasks like transferring binaries to the NAS or interacting with the local DSM Web API. No patterns indicating exfiltration of sensitive information to external, untrusted domains were observed.
[CREDENTIALS_UNSAFE]: The skill mentions sensitive file paths such as ~/.ssh/authorized_keys and .env equivalents in the context of diagnostic hygiene and security hardening (e.g., recommending 2FA and disabling default accounts). No hardcoded credentials or private keys are included in the skill files.
[INDIRECT_PROMPT_INJECTION]: The skill interacts with system files and command outputs from the Synology appliance. 1. Ingestion points: Commands like 'cat /etc.defaults/VERSION' and 'synopkg list --name' bring external system data into the agent's context. 2. Boundary markers: The skill defines an 'Output Contract' and strict 'Rules' for verification, which serve as logical boundaries. 3. Capability inventory: The agent is empowered to execute system-level commands (sudo, mdadm, btrfs) and network operations (scp, curl). 4. Sanitization: The skill emphasizes 'Step 1: Identify the appliance' and 'Step 5: Verify against the runtime' to validate system state before acting on potentially deceptive data.

synology-dsm