The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted output (journal logs, agent status) and explicitly implements a security control to mitigate injection. It uses a nonce-delimited <untrusted-data> wrapper defined in references/untrusted-output.md to ensure that workload-emitted data is treated as non-authoritative data only, preventing instructions in logs from influencing the agent's behavior.
[DATA_EXFILTRATION]: The skill handles sensitive configuration files (e.g., /etc/observe-agent/observe-agent.yaml) that contain ingest tokens. It implements explicit redaction logic using sed to remove the token from the file content before it is passed to the agent context. Additionally, it provides a safe 'fingerprinting' method using Python to verify token properties (length and suffix) without exposing the secret value.
[COMMAND_EXECUTION]: The skill generates shell commands for system troubleshooting (e.g., journalctl, systemctl, apt, yum). These commands are designed to be run by the user on the target host, not executed directly by the agent shell. Destructive commands, such as removing package configuration or backups, are gated by explicit mandatory user confirmation steps.
[PRIVILEGE_ESCALATION]: The skill uses sudo for administrative tasks like restarting services or reading system logs. This is appropriate for the stated purpose of host-level agent debugging and follows standard Linux administration patterns.
[EXTERNAL_DOWNLOADS]: The skill references official Observe documentation and facilitates agent reinstallation via standard package managers (apt, yum) from established repositories, which is considered safe behavior for a maintenance tool.

debug-linux-host-collection