The Agent Skills Directory

[COMMAND_EXECUTION]: The framework provides an ExecuteShellCommandTool enabling the agent to run arbitrary shell commands on the host system.
[COMMAND_EXECUTION]: The BraveModeConfirmationHandler allows for the automatic execution of shell commands without human oversight, creating a significant risk of unauthorized system operations.
[DATA_EXFILTRATION]: The skill includes ReadFileTool, ListDirectoryTool, and WriteFileTool, which grant the agent the ability to read and modify sensitive files on the local file system.
[DATA_EXFILTRATION]: The TraceFeatureMessageRemoteWriter allows agent execution data and traces to be sent to arbitrary remote HTTP endpoints, which could be used to exfiltrate system information or session data.
[REMOTE_CODE_EXECUTION]: The combination of the WriteFileTool (to create scripts) and the ExecuteShellCommandTool (to run them) provides a direct vector for remote code execution if the agent is compromised.
[PROMPT_INJECTION]: As an LLM-orchestration framework, it is susceptible to prompt injection attacks where malicious instructions could trigger the dangerous file and shell capabilities.

koog