As AI model capabilities are enhanced, so are the demands of business clients for data privacy and abuse monitoring. On August 19, OpenAI previewed a new service called Private Safety Production, stating that the system automatically recognizes potential misuse without retaining client data.
Aim for business privacy concerns
This service is now open to some clients. OpenAI states that the new system is monitored by an automated agent and does not require a manual article-by-article view of the content of the client dialogue. Its objective is to meet two sets of needs simultaneously: first, to detect malicious use and, second, to minimize the retention of enterprise-sensitive data on the platform.
The article mentions that OpenAI's existing API has adopted the Zero Data Retenance model. This model allows for the detection of unusual use by single session, but does not keep customer data for long periods. The new programme expands the scope of surveillance on this basis.
Unusual recognition of multiple sessions
Unlike one-way conversations, Private Safety Production is focused on long-term behavioural monitoring of cross-conferences. OpenAI states that some abusers may break requests into multiple conversations to reduce the probability of detection, such as step-by-step seeking malicious software or cyberattack-related help.
In such cases, the system automatically analyses multiple rounds of input and output to determine whether there are signs of potential non-compliance. OpenAI states that if the system is triggered, it will only send a " narrower signal" to the company, indicating a specific type of activity, rather than directly retrieving the full content of the dialogue.
If OpenAI considers that further processing is required, it will contact the client to understand the background or to work with the client. It is up to the client to decide whether to provide more data to OpenAI.
Compared to the Anthropic policy
This move is in stark contrast to the recent data retention policy of Anthropic. According to reports, in July Anthropic announced that a partial “covered model” could retain user data for 30 days, including relevant sessions and their contents, for security clearance.
This arrangement has led to dissatisfaction among a number of business clients, particularly companies that handle a large amount of sensitive information. They fear that the data will be kept by AI or will be viewed in the context of security reviews. It was also mentioned that in most cases, Anthropic also uses zero data retention, with some exceptions for high-capacity models.
In terms of industry competition, enterprise-level privacy and safety capabilities are becoming important outlets for AI companies to compete for customers. According to TechCrunch, the introduction of a new programme at this time is also a response to the policy controversy of Anthropic in the business services market and seeks to increase the number of clients with stringent data-processing requirements.