OpenAI has begun testing a new security mechanism for customers who access its flagship artificial intelligence models through paid subscriptions. According to news BloombergThese measures are designed to enhance protection in an environment where customers are increasingly trusting artificial intelligence to perform complex tasks and handle sensitive information.

Image source: OpenAI (added)
Currently, OpenAI guarantees a “zero data retention” policy to some customers using the API, that is, requests and responses will not be stored after AI is completed. Now, by testing a mechanism called “private security processing,” OpenAI aims to teach the system to recognize a set of patterns rather than a single request and response. This will help prevent hacking attempts by attackers and AI agents whose behavior deviates from specified security standards.
The company explained the need for this change as modern AI models are increasingly used to perform complex, multi-stage tasks, while the role of tools that can operate without detailed user instructions is growing. Aleah Houze, product policy manager at OpenAI, said it’s not always possible to detect potential risks by just looking at a pair of requests and responses, because suspicious sequences only emerge when multiple interactions over time are analyzed. She argued that individual calls to artificial intelligence may look like perfectly legitimate research in cybersecurity, but that taken together they could indicate preparation for an attack.
Howes gave the example of a user who discovered information about certain software vulnerabilities in a conversation and subsequently became interested in remote access capabilities and methods of detecting such threats. The requests themselves may not raise suspicion, but their connection to each other can prove malicious behavior.
OpenAI anticipates that private security processing will allow it to take such context into account, thereby more effectively combating attackers and potentially dangerous AI agents, while keeping OpenAI itself unable to access the content requested by the client. Testing is conducted with customers such as Microsoft and Databricks. The launch of the system and the release of a report describing the technology are planned for September.
If you find an error, select it with your mouse and press CTRL+ENTER.
