FIX: support moderation errors for MAI image models (#1890)

Author: fdubut

pyrit/prompt_target/openai/openai_error_handling.py

@@ -75,7 +75,7 @@ def _is_content_filter_error(data: Union[dict[str, object], str]) -> bool:
         # Check for explicit content_filter or moderation_blocked codes
         error_obj = data.get("error")
         code = error_obj.get("code") if isinstance(error_obj, dict) else None  # type: ignore[ty:invalid-argument-type]
-        if code in ["content_filter", "moderation_blocked"]:
+        if code in ["content_filter", "content_safety_violation", "moderation_blocked"]:
             return True
         # OpenAI uses "invalid_prompt" for model-level safety blocks (e.g. CBRN topics).
         # Only treat it as a content filter when the message indicates a safety block,

pyrit/prompt_target/openai/openai_image_target.py

@@ -299,7 +299,8 @@ async def _send_edit_request_async(self, message: Message) -> Message:
         # Construct request parameters for image editing
         image_edit_args: dict[str, Any] = {
             "model": self._model_name,
-            # Single image can be sent (and MUST BE in MAI models) as a tuple, multiple images must be a list
+            # Single image sent as a tuple (mandatory for targets that support only one image input such as MAI,
+            # also supported by other targets such as OpenAI). Multiple images always sent as a list.
             "image": image_files[0] if len(image_files) == 1 else image_files,
             "prompt": text_prompt,
             "size": self.image_size,