Audit a WAF 403 Dump Before a Model Sees the Cookie You Blocked

Описание и применение

Инженеры случайно вставляют полные логи WAF 403, содержащие чувствительные заголовки, в модели ИИ, раскрывая сессии. Решение: сканер, который отклоняет логи с Cookie, Authorization и другими чувствительными данными, и использовать только редактированные метаданные.

Задача

Engineers inadvertently paste full WAF 403 logs containing sensitive headers (Cookie, Authorization, etc.) into AI models, leaking session credentials and violating trust boundaries.

Результат

By rejecting logs that contain sensitive headers or tokens, the organization prevents accidental credential leakage into AI models, maintaining trust boundaries and reducing the risk of credential compromise.

Рабочий процесс

• 1. Receive WAF 403 log in incident channel.
• 2. Copy raw log line into AI prompt.
• 3. AI model answers why the request was blocked.
• 4. Sensitive header values (Cookie, Authorization, etc.) are exposed to the model.
• 5. No control prevents leakage.
• 6. Implement a prompt-boundary scanner that rejects logs containing sensitive headers or tokens.
• 7. Use redacted fixtures to test the scanner.
• 8. Replace raw logs with redacted metadata before prompting the model.

Доказательства

Never send these in a why-was-this-blocked prompt: Cookie, Set-Cookie, Authorization, CSRF tokens, internal IPs, JWT strings.

Источник

Build a prompt-boundary fixture… A command that exits 1.

Источник