The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources (API responses for Agents, Sessions, and Events) and possesses powerful capabilities (command execution via arkcli and file writing).
Ingestion points: Data enters the agent context through the execution of arkcli commands that fetch remote resource configurations, session history, and skill metadata (found in SKILL.md, references/agent.md, and references/skills.md).
Boundary markers: The skill does not define specific delimiters to separate untrusted API data from internal instructions, relying on the agent's ability to parse JSON/YAML structures.
Capability inventory: The skill allows for local command execution (arkcli), file uploads (agent file upload), and sending messages to other agents (events send).
Sanitization: There are no explicit instructions for sanitizing or escaping the content of external API responses before processing or re-interpolating them into downstream tasks.
[COMMAND_EXECUTION]: The skill's primary function is to execute the arkcli binary on the host system to manage cloud resources. While this is the intended purpose, it represents a significant capability that could be misused if the agent is compromised by external data.
[REMOTE_CODE_EXECUTION]: The skill provides instructions for initializing remote environments with custom scripts (--setup-script @./bootstrap.sh in references/session-files.md). Although this executes in the managed cloud environment rather than the agent's host, it is a mechanism for code deployment and execution managed by the agent.

arkcli-agent