OpenAI Answers Anthropic with a Monitoring Service That Stores No Data

The developers of large language models find themselves in an awkward spot: the more powerful the systems become, the louder the demand to monitor how they are used—and the sharper the conflict between that oversight and the privacy of enterprise clients. OpenAI has decided to turn this tension into a competitive advantage.

The company has announced a preview version of a service called Private Safety Processing for select customers. It's an automated system that tracks possible abuse while retaining not a single byte of user data. The jab is aimed at Anthropic, whose July retention policy has ruffled some customers: the lab keeps all sessions and conversations for 30 days—though only for “covered models,” a category that includes Mythos-class systems and “future models with similar capabilities.”

From the Session to the Long Haul

Most AI developers, including OpenAI itself, already adhere to the principle of Zero Data Retention: agents inside the API watch for violations within a single session, without storing data or involving humans. Anthropic largely honors ZDR too—except for those very “covered models” such as Fable.

Private Safety Processing, according to OpenAI, extends this approach. It's a matter of “long-term” monitoring that analyzes the inputs and outputs of not one conversation but several at once. The logic is simple: a bad actor—say, someone trying to assemble malicious code for a cyberattack—can spread their requests across different sessions to avoid getting caught. The new agent is capable of correlating these conversations and recognizing the signs of abuse without showing the exchanges to any human employees.

If the system trips, it sends OpenAI a “narrowly defined signal” about a specific type of activity. On that basis, the company decides whether action is needed. If necessary, OpenAI contacts the customer for clarification, and the customer then decides for themselves whether to share the data.

By way of contrast, OpenAI points out how the process works at its competitor: Anthropic does allow humans to review data, but only through a “controlled access channel” with a narrow set of approved reviewers, and each such session is logged in a tamper-proof record that reviewers cannot erase or alter.

The Stakes Are High

Behind the technical details lies a fierce rivalry between the two labs. According to recent figures, OpenAI's revenue grew more slowly in the second quarter than Anthropic's, whose annualized revenue rate is already pegged at $65 billion. Anthropic's investors are talking about a possible IPO at a $2 trillion valuation, and OpenAI is preparing its own market debut as well. In this climate, customer privacy becomes not only an ethical question but also a battlefield in the fight for the enterprise client.

Source: TechCrunch