The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The scripts ssh-download.sh, ssh-exec.sh, ssh-job.sh, ssh-test.sh, and ssh-upload.sh all utilize sshpass -p "$PASSWORD" to handle authentication. This method is insecure as it exposes the password in the process table, making it readable by other users on the local system.
[COMMAND_EXECUTION]: The skill facilitates the execution of arbitrary commands on remote hosts and within containers via ssh, docker exec, and kubectl exec. This provides a significant attack surface if the agent is manipulated into running malicious commands.
[DYNAMIC_EXECUTION]: Multiple scripts use eval to execute shell commands built from interpolated variables. For instance, ssh-download.sh and ssh-upload.sh use eval to run constructed rsync or scp commands. Additionally, ssh-job.sh executes remote commands by decoding Base64 strings directly into bash -c. These patterns are susceptible to command injection if the configuration data (profiles) or inputs are improperly sanitized.
[DATA_EXFILTRATION]: The inclusion of ssh-download.sh allows the agent to retrieve files and directories from remote servers or containers to the local system, which could be used to exfiltrate sensitive data from remote environments.
[OBFUSCATION]: The ssh-job.sh script encodes remote commands using Base64 before transmission. While the skill claims this is for handling shell quoting issues, it serves to obfuscate the intent and content of commands from simple text-based monitoring.
[PERSISTENCE]: The ssh-job.sh script utilizes nohup and background processes (&) to ensure that tasks continue to run on the remote server after the initiating session has ended, which can be exploited for maintaining persistence.

ssh-dev-suite/connect