This PR adds image moderation via Azure AI Content Safety under an experiment flag 'ai-content-safety'.

PR updates include:

• Added AzureAiContentSafety service to backend and associated key in secrets (azure_ai_content_safety_key)
• Added call to AzureAIContentSafety's moderate_image in ImageModeration's moderation_image.
• Added new backend endpoint in files_api.rb: "POST /v3/images/moderate-ai-content-safety"
• Removed Firehose logging to 'azure-content-moderation' and now unused image_url param. See https://codedotorg.atlassian.net/browse/INF-1996
• Added an experiment flag 'ai-content-safety', and if flag is turned on,moderateImage uses Azure AI Content Safety. The default service is still Azure Content Moderator.

Note that I added to azure_ai_content_safety_key to AWS secrets using bin/update_secrets which was cool and much easier than adding a secret manually (which I've done in the past).

I added the severity level blocks (currently at 2 for each category) on the front-end. There has been discussion about folks wanting to know why an image is blocked - eee Slack thread. Follow-up can include adding metadata to when an image is flagged with the categories and severity level.
FYI, we can test images on https://contentsafety.cognitive.azure.com/image! So if product/curriculum decide on different severity blocked levels for different categories, we can update.

Links

• Jira: https://codedotorg.atlassian.net/browse/SL-1633

Testing story

Tested locally and logged the json returned (in dev console) when experiment flag was turned on.

In Web Lab 2 - 'ok':

In Web Lab 2 - 'flagged':

In Sprite Lab - 'ok':

In AI Chat Lab - 'ok':

Follow-up

• Add categories and severity level in Statsig reporting if image flagged by AI Content Safety.
• Once tested on prod with flag, can remove experiment flag and migrate to AI Content Safety.
• Remove Azure Content Moderator and associated code.
• Consider moving the scaling of small images to backend.

Deployment notes

Privacy and security