arkcli-chat
Pass
Audited by Gen Agent Trust Hub on Sep 21, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using the vendor's 'arkcli' binary and standard system utilities like 'jq' and 'mktemp' for output parsing and temporary file management.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multimodal inputs and model responses, creating a surface for indirect prompt injection.
- Ingestion points: Processes local files provided via the @path syntax (images, video, audio) and ingests responses from the model API.
- Boundary markers: Employs a 'strict' mode for JSON Schema validation (--text-strict) which provides structural boundaries for model outputs.
- Capability inventory: Includes shell execution, file system access (reading inputs, writing results), and network API communication.
- Sanitization: Uses client-side JSON Schema validation to ensure the integrity and format of the data before it is delivered to the user or saved to disk.
- [DATA_EXFILTRATION]: The skill performs network operations via the 'arkcli' tool to Volcengine's Ark platform infrastructure (Responses API). The instructions include explicit security practices, such as preventing the logging or echoing of API keys, to protect user credentials during these operations.
Audit Metadata