Toxicity in Unmask API

How toxicity analysis appears in Unmask API responses — auditing restored content for safety compliance.

{
  "unmask": [
    {
      "value": "<PER>hSw8kAEB10</PER> lives in <ADDRESS>748785848000</ADDRESS>"
    }
  ]
}

curl -X PUT https://protecto-trial.protecto.ai/api/vault/unmask \
  -H "Authorization: Bearer <AUTH_TOKEN>" \
  -H "Content-Type: application/json; charset=utf-8" \
  -d '{
    "unmask": [
      {
        "value": "<PER>hSw8kAEB10</PER> lives in <ADDRESS>748785848000</ADDRESS>"
      }
    ]
  }'

{
  "data": [
    {
      "value": "George Williams lives in Washington",
      "token_value": "<PER>hSw8kAEB10</PER> lives in <ADDRESS>748785848000</ADDRESS>",
      "toxicity_analysis": {
        "toxicity": 0.0008883,
        "severe_toxicity": 0.0001045,
        "obscene": 0.0001825,
        "threat": 0.0001108,
        "insult": 0.0001754,
        "identity_attack": 0.0001380
      }
    }
  ],
  "success": true,
  "error": { "message": "" }
}

The Unmask API also returns toxicity analysis for the unmasked text. This allows you to audit restored content, apply moderation after de-tokenization, and enforce stricter controls on data access workflows.

How it appears in the response

Important notes

policy_name is optional in unmask requests
Toxicity scores are always returned when unmasking succeeds
Unmask authorization rules still apply — toxicity detection does not bypass access controls
Scores reflect the restored original text, not the tokenized form

Use cases for unmask toxicity

Use case	Description
Post-unmask moderation	Check restored content before displaying to users
Audit logging	Log toxicity scores alongside all unmask operations
Access control augmentation	Apply additional review when high toxicity + unmask are combined
Compliance reporting	Record safety signals for regulated workflows

Combining identity_attack scores with unmask audit logs is useful for detecting whether sensitive personal data is being accessed in discriminatory contexts.

Was this page helpful?

Last updated Apr 6, 2026

Built with Documentation.AI