The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes command templates that interpolate unvalidated user-provided input, such as project paths and tool arguments, into shell commands executed on a remote host.
Evidence: The command templates for running EDA tools and monitoring background processes via ssh -o BatchMode=yes vmhw in SKILL.md.
[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to read and interpret log files and process metadata from the target environment to verify task success, which constitutes a vulnerability surface if those files contain adversarial content.
Ingestion points: Data is ingested from the target VM via tail -n 60 on log files and pgrep -fl for process lists as described in SKILL.md.
Boundary markers: Absent. No delimiters or warnings are provided to the agent to disregard potentially malicious instructions within the logs.
Capability inventory: The skill provides the agent with extensive remote execution capabilities via SSH, including file system access and GUI interaction tools.
Sanitization: No sanitization or validation logic is specified for the external data ingested from logs or process outputs before interpretation by the agent.

vmhw