update language

Author: PatrickFarley

articles/ai-foundry/responsible-ai/openai/transparency-note.md

@@ -607,7 +607,7 @@ Azure OpenAI service can support a wide range of applications like search, class
 - **Limit the length, structure, and rate of inputs and outputs.**  Restricting the length or structure of inputs and outputs can increase the likelihood that the application will stay on task and mitigate, at least in part, any potentially unfair, unreliable, or offensive behavior. Other options to reduce the risk of misuse include (i) restricting the source of inputs (for example, limiting inputs to a particular domain or to authenticated users rather than being open to anyone on the internet) and (ii) implementing usage rate limits. 
 - **Encourage human review of outputs prior to publication or dissemination.** With generative AI, there is potential for generating content that might be offensive or not related to the task at hand, even with mitigations in place. To ensure that the generated output meets the task of the user, consider building ways to remind users to review their outputs for quality prior to sharing widely. This practice can reduce many different harms, including offensive material, disinformation, and more. 
 - **Implement additional scenario-specific mitigations.**  Refer to the mitigations outlined in [Evaluating and integrating Azure OpenAI for your use](#evaluating-and-integrating-azure-openai-natural-language-and-vision-models-for-your-use) including content moderation strategies. These recommendations do not represent every mitigation required for your application. Newer models such as GPT-4o and reasoning models may provide responses in sensitive scenarios and are more likely to attempt to reduce potentially harmful outputs in their responses rather than refuse to respond altogether. It's important to understand this behavior when evaluating and integrating content moderation for your use case; adjustments to filtering severity may be needed depending on your use case.
-- **Avoid triggering mandatory safeguards.** Azure Direct Models may have safeguards to prevent security exploits including output of raw CoT. Use of a model in a manner that creates a security exploit, including attempting to extract the raw CoT, for example, by circumventing these safeguards, violates the Acceptable Use Policy for Online Services and may result in access to the service being limited. Some models have safeguards specific to certain risk types, for example, biosecurity. For greater detail on best practices, see the [OpenAI o1 System Card](https://cdn.openai.com/o1-system-card-20241205.pdf), [o3-mini System Card](https://openai.com/index/o3-mini-system-card/), [o3/o4-mini System Card](https://openai.com/index/o3-o4-mini-system-card/), and [GPT-5 System Card](tbd).
+- **Avoid triggering mandatory safeguards.** Azure Direct Models may have safeguards to prevent security exploits including output of raw CoT and biosecurity content. Use of a model in a manner that creates a security exploit or evades or attempts to evade a protection on the model,including by circumventing these safeguards, violates the Acceptable Use Policy for Online Services and may result in access to the service being limited. For greater detail on best practices, see the [OpenAI o1 System Card](https://cdn.openai.com/o1-system-card-20241205.pdf), [o3-mini System Card](https://openai.com/index/o3-mini-system-card/), [o3/o4-mini System Card](https://openai.com/index/o3-o4-mini-system-card/), and [GPT-5 System Card](tbd).
 
 #### Best practices and recommendations for fine tuning