Sec-Deserialization-Gadget-Risk

The model receives the prompt (and optional system message). The run uses scorer rubric_json_metrics with the JSON configuration below. Pass/fail and partial credit are determined entirely by that scorer against the model output; no human grading.

Return JSON only with keys verdict, risk, mitigation. A worker deserializes untrusted job payloads with a language serializer that can instantiate arbitrary classes.

{
  "metrics": {
    "correct": {
      "checks": [
        {
          "contains": [
            "unsafe deserialization"
          ]
        },
        {
          "contains": [
            "untrusted"
          ]
        },
        {
          "contains": [
            "arbitrary classes"
          ]
        }
      ]
    },
    "hidden": {
      "checks": [
        {
          "contains": [
            "JSON schema"
          ]
        },
        {
          "contains": [
            "allowlist"
          ]
        },
        {
          "contains": [
            "code execution"
          ]
        }
      ]
    }
  }
}