The Agent Skills Directory

[SAFE]: The skill consists entirely of documentation and architectural guidance. No scripts, binaries, or automated tasks were found in the skill body or referenced files.\n- [SAFE]: The instructions emphasize security principles such as isolation, least privilege (selecting the smallest environment matching the task needs), and strict network policies.\n- [SAFE]: There are no hardcoded secrets or sensitive file paths accessed within the skill definition. The mention of a 'secret interface' refers to the configuration of the environment being provisioned, not an exposure of host credentials.\n- [SAFE]: The skill instructs the agent to use official provider documentation and native SDKs/CLIs for provisioning, which minimizes the risk of using untrusted or third-party components.\n- [PROMPT_INJECTION]: The skill addresses potential indirect injection risks by providing an evidence chain for secure workspace management: 1. Ingestion points: User requests for sandbox or VM provisioning. 2. Boundary markers: Mandatory identity and isolation boundaries, resource lifetime limits, and the use of filesystem canaries. 3. Capability inventory: Provider-native SDK/CLI execution, filesystem operations, and command execution for verification. 4. Sanitization: Relies on provider-native interfaces and matching isolation levels to data sensitivity. Findings are categorized as safe due to the implementation of these defensive instructions.

provision-agent-workspace