The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on shell-based interactions with the local system to function. It uses docker ps, docker inspect, and docker logs to retrieve container metadata and stream execution output. It additionally uses jq to extract values from training manifest files and tail to display progress. These operations incorporate variables such as container names and directory paths sourced from user input or file content.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data by reading training manifests (run.json), workflow logs (*.log), and evaluation results (test_result.csv). This content is then displayed to the agent, creating a surface for potential indirect prompt injection if these files were to be modified by an adversary to include malicious instructions.
[REMOTE_CODE_EXECUTION]: In its status summary, the skill surfaces a 'Replay command' extracted from the cmd_replay field within the run manifest. This pattern encourages the execution of arbitrary command strings stored in a configuration file, which is a common utility for workflow management but represents a potential vector if the manifest file is untrusted.

kermt-monitor