kitaru-adapter-builder
Pass
Audited by Gen Agent Trust Hub on Sep 7, 2026
Risk Level: SAFE
Full Analysis
- [DATA_EXPOSURE_AND_EXFILTRATION]: The skill proactively mitigates data exposure by instructing the agent to always redact or omit authorization headers, cookies, API keys, passwords, and other credentials from recorded data and logs (seen in
references/python-adapters.mdandreferences/typescript-adapters.md). - [UNVERIFIABLE_DEPENDENCIES]: The skill mandates explicit user approval before installing or changing any dependencies, specifying the exact package, version, and command. It also emphasizes using official, published versions of Kitaru-related packages (e.g.,
@zenml-io/kitaru) rather than unverified sources. - [COMMAND_EXECUTION]: The skill limits command execution to safe project inspection and requires approval for write-capable tools. It specifically warns against using improvised event-loop bridges or private framework imports that could lead to unstable or insecure execution environments.
- [INDIRECT_PROMPT_INJECTION]: While the skill processes project files (untrusted data), it implements a robust 'Core contract' and validation methodology that focuses on read-only inspection, boundary markers, and allowlisted data projection to prevent injected instructions from influencing agent behavior.
- [DYNAMIC_EXECUTION]: The skill explicitly discourages unsafe dynamic practices, such as handwritten REST calls to replace SDK methods or unsafe deserialization, instructing the agent to reject dangerous keys and functions in serialized payloads.
Audit Metadata