The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill documentation and provided examples (SKILL.md, references/example-tasks.md) recommend a high-risk pattern for installing dependencies in verification scripts by piping a remote script directly to a shell.
Evidence: The command curl -LsSf https://astral.sh/uv/0.9.7/install.sh | sh is suggested for installing the 'uv' package manager. This pattern executes untrusted code from a remote source with the privileges of the executing user.
[COMMAND_EXECUTION]: The Harbor runtime model described in the skill involves the automated execution of shell scripts with execution permissions within a containerized environment.
Evidence: The skill documents that Harbor automatically runs chmod +x /tests/test.sh and then executes the script during the verification phase. This provides a vector for arbitrary command execution if the script content is influenced by untrusted task definitions.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to facilitate the creation of task environments and instructions for other AI agents, creating a surface for indirect injection attacks.
Ingestion points: instruction.md files, which are the primary input for agents in the Harbor evaluation framework.
Boundary markers: While the skill documents structured task components, it lacks explicit guidance on using boundary markers or delimiters to isolate untrusted content within instructions.
Capability inventory: The skill documents that tasks have significant capabilities, including Docker container access, shell execution for tests, and potential network access, which could be abused if an agent is misled by injected instructions.
Sanitization: No specific sanitization or validation mechanisms are recommended for the data interpolated into task instructions.

harbor-task-creator