codebolt-agent-development
Pass
Audited by Gen Agent Trust Hub on Mar 14, 2026
Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The documentation describes APIs and modifiers for executing arbitrary shell commands through the terminal.
- Evidence: Mentions of
codebolt.terminal.executeCommandandShellProcessorModifierinSKILL.mdandreferences/processors.md. - [REMOTE_CODE_EXECUTION]: The skill includes code snippets demonstrating the use of
eval()for dynamic data processing within custom tool implementations. - Evidence: Example implementation of a calculator tool using
eval()inreferences/level2-base-components.mdandreferences/processors.md. - [PROMPT_INJECTION]: The agent architecture described involves processing untrusted data which creates an indirect prompt injection surface.
- Ingestion points:
codebolt.onMessageandcodebolt.fs.readFile(SKILL.md, references/level1-direct-apis.md). - Capability inventory: Commands for terminal execution, browser control, and file modifications are available across the SDK.
- Boundary markers: Documentation specifies that tool approval and workspace boundaries are managed at the application level rather than within the agent code.
- Sanitization: Provided code examples show direct interpolation of retrieved data into prompts without explicit sanitization or filtering logic.
Audit Metadata