Content filter and policy violation issues when using openAI API

syyu 5

Too frequent rule violations occur when using openAI services.

Even prompts that are nothing special, such as “Make it like the picture” or “Make it more exciting,” do not receive answers because against the rules.

Do you know a solution to this?

The following error occurs:

openai.BadRequestError: Error code: 400 - {'error': {'code': 'content_policy_violation', 'inner_error': {'code': 'ResponsibleAIPolicyViolation', 'content_filter_results': {'jailbrea

k': {'detected': False, 'filtered': False}}}, 'message': 'Your request was rejected as a result of our safety system. Your prompt may contain text that is not allowed by our safety system

.', 'type': 'invalid_request_error'}}

openai.BadRequestError: Error code: 400 - {'error': {'message': "The response was filtered due to the prompt triggering Azure OpenAI's content management policy. Please modify your prompt and retry. To learn more about our content filtering policies please read our documentation: https://go.microsoft.com/fwlink/?linkid=2198766", 'type': None, 'param': 'prompt', 'cod e': 'content_filter', 'status': 400, 'innererror': {'code': 'ResponsibleAIPolicyViolation', 'content_filter_result': {'hate': {'filtered': False, 'severity': 'safe'}, 'self_harm': {'filte red': False, 'severity': 'safe'}, 'sexual': {'filtered': True, 'severity': 'medium'}, 'violence': {'filtered': False, 'severity': 'safe'}}}}}

Axel B Andersen 146 Reputation points

2024-09-11T10:57:03.6866667+00:00

Since yesterday we started getting this exact same issue on one of our installations - attempt to jailbreak. On another installation we are constantly blocked due to Violence. And that is when just sending Good morning. It seems like the content filters on some endpoints is going crazy. All our deployments are in sweden central and only a single deployment has the error todya, but that is on every single query.
YutongTie-MSFT 53,211 Reputation points

2024-09-11T17:36:25.2833333+00:00

Hello @syyu @Axel B Andersen Thanks for reaching out to us and sorry for the late response, could you please share the region you are working on so that we investigate it?

Regards,

Yutong
Axel B Andersen 146 Reputation points

2024-09-11T17:42:22.3366667+00:00

Thanks for looking into this, Yutong! Our deployments are in Sweden Central. Any instruction larger then "you are a helpful assistant" trigger either self-harm, violence or jailbreak.
YutongTie-MSFT 53,211 Reputation points

2024-09-11T18:40:51.54+00:00

Hello @Axel B Andersen Could you please also share the model you are working? If you are not working in a Azure OpenAI Studio, please specify which API version you are working on. I can not reproduce the same issue in model gpt 4o and gpt 4 turbo 32.

Could you please also confirm you have no content filter or blocklist applied on ?

Thanks!
Axel B Andersen 146 Reputation points

2024-09-11T18:51:34.3266667+00:00

We use semantic kernel and I think it uses the april preview api version. We have at least 10 of these deployments in Sweden Central, but only 2 of them have this error. It started yesterday and no changes have been made to kernel version or prompts.
YutongTie-MSFT 53,211 Reputation points

2024-09-11T19:07:15.89+00:00

Hello @Axel B Andersen

Thanks for the response, it seems there is a custom content filter, but if as you said there is nothing chance before the error happened, I would suggest you raise a support ticket for this case so that a support engineer can help you investigate the root cause. I am happy to enable you a free ticket if you have no support plan, please let me know.

Just want to double check @syyu, please let me know your scenario if the case is not the same one as Axel.

Regards,

Yutong
Axel B Andersen 146 Reputation points

2024-09-11T19:23:30.32+00:00

Thanks, Yutong. I will create a support case in the morning. The custom content filter has been added today with everything set to low/off in a try to mitigate the issue. It seems to have had some effect on one of the installations.
YutongTie-MSFT 53,211 Reputation points

2024-09-11T19:29:03.1166667+00:00

Do you need me to enable a free ticket or you have the support plan? @Axel B Andersen

Please share the support ticket number to me so that I can track as well. Thanks!
Axel B Andersen 146 Reputation points

2024-09-11T19:32:13.98+00:00

We will most likely be able to create a support case on the affested installation. I will reach out if not snd I will keep you posted. Thanks 😊
YutongTie-MSFT 53,211 Reputation points

2024-09-12T04:43:27.12+00:00

Hello @syyu @Axel B Andersen

Thanks for reaching out to us again, we have escalated this issue to product team and product team acknowledged this issue. The next revision is 10/15, so at that point we should expect this to be resolved. I hope this helps!

Regards,

Yutong

1 answer

Axel B Andersen 146 Reputation points

2024-09-20T10:07:41.5233333+00:00

This issue has been resolved. Turns out that 2024-05-13 sometimes omits the content parameter of the message. Updating to 2024-08-06 removed that issue. Additionally, the filter still evaluated, that the initial prompt was trying to jailbreak the model. Changing the prompt removed the jailbreak issue.

The prompt that generated a jailbreak warning one deployment did not generate a jailbreak warning on another deployment. It seems to be important to test all prompts on the specific deployments they should run on.
Please sign in to rate this answer.

0 comments No comments
Sign in to comment

Use comments to ask for clarification, additional information, or improvements to the question.

Share via

Content filter and policy violation issues when using openAI API

1 answer

Your answer