The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The SKILL.md file contains instructions directing the agent to retry commands outside the sandbox environment if they fail. Specifically, it states that if an operation fails while sandboxed, the agent should retry the command unsandboxed to access the host credential store and request privilege escalation if required.
[PROMPT_INJECTION]: The workflow-vibe.md prompt reference instructs the agent to adopt an "Action-First Mindset," where it is told to execute technical wiring and code modifications immediately without describing the changes or asking for user permission. This instruction explicitly forbids the agent from seeking confirmation (labeling it a "violation"), which bypasses standard human-in-the-loop oversight.
[COMMAND_EXECUTION]: The skill relies on a bundled JavaScript CLI tool located at scripts/aistudio.js. The agent is instructed to run this tool frequently via node for core management of AI Studio artifacts, including creating, modifying, and validating project files.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to build applications that ingest data from external sources such as Business Objects and REST APIs (as documented in app-best-practices.md and app-ingestion.md). While the skill provides instructions on how to mitigate injection risks in the workflows it generates (e.g., using boundary markers and system guards), the agent itself acts as a processing engine for untrusted external data with the capability to execute shell commands via the CLI script.

aistudio