Privacy Wars: OpenAI Escalates Battle with Anthropic!
OpenAI has introduced Private Safety Processing, a new privacy-centric system designed to monitor AI misuse across multiple sessions without retaining customer data. This approach contrasts sharply with rival Anthropic's data-retention policy for "covered models," where user data is kept for 30 days. The move highlights the intensifying corporate competition and the delicate balance between AI safety and user privacy.
As artificial intelligence models continue to advance in power and capability, the potential for their misuse has grown significantly, fueling an increasing demand for robust safety guardrails. AI companies are now navigating a delicate balance, aiming to respect their enterprise customers’ privacy while rigorously monitoring for possible issues and abuse. Recognizing this crucial challenge and sensing an opportunity to differentiate itself from rival Anthropic, OpenAI has announced a new privacy-centric safety approach specifically designed for misuse monitoring.
OpenAI is currently previewing a new service to select customers, which it calls Private Safety Processing. This innovative system operates as an automated mechanism that vigilantly watches for potential abuse without retaining any of the customer’s data. This methodology directly contrasts with Anthropic’s recently announced data-retention policy. Anthropic’s policy, which has caused concern among some customers, allows the AI lab to keep user data—including all sessions and the conversations within them—for a period of 30 days. This applies to what Anthropic terms “covered models,” which encompass all Mythos-class models and “future models with similar capabilities,” as stated by the company upon its announcement in July. While designed for safety purposes to allow the lab to sift and analyze potential impropriety, this policy has deeply troubled enterprises that manage large volumes of sensitive data and are hesitant to have it harbored or inspected by the AI lab.
It is important to note that OpenAI, along with most other AI companies, already provides customers with a relative level of privacy by adhering to a policy known as Zero Data Retention (ZDR). ZDR employs agents within the OpenAI API to monitor for abuse on a per-session basis, thereby ensuring customer data is not retained by the company while still enabling scans for illicit activity without requiring human intervention. Anthropic also largely abides by ZDR, with the specific exception of its “covered models” like Fable, where the data-retention policy applies.
OpenAI describes Private Safety Processing as a new technology that significantly widens the scope of ZDR. It functions as a form of long-horizon safety monitoring, assessing the inputs and outputs across multiple conversations, rather than being confined to single sessions. The monitoring is conducted by an automated agent which, if triggered, captures interactions and analyzes them across sessions for signs of potential misuse. This new technology is crucial for OpenAI to detect malicious use of AI that unfolds over multiple sessions, a spokesperson informed TechCrunch. For instance, a bad actor attempting to engineer malware for a cyberattack might deliberately spread out their requests to evade immediate detection. Private Safety Processing can analyze these multiple conversations for signs of abuse without requiring human review of a user’s discussions.
In the event that the system is triggered, it transmits a “narrowly defined signal” to OpenAI, which alerts the company to a specific type of activity. Based on this signal, OpenAI can then decide if “enforcement is necessary.” Should enforcement be deemed appropriate, OpenAI will reach out to the customer for additional context or to collaborate on resolving the issue, with the customer retaining the discretion to share data with OpenAI at their own choosing, as clarified by the spokesperson. In contrast, Anthropic’s policy states that human review of customer data can occur, but only through a “controlled access path” that involves “a small set of approved reviewers.” Furthermore, every one of these review sessions is “recorded in a tamper-proof log that reviewers cannot suppress or modify,” according to the company.
The corporate competition between OpenAI and Anthropic is notably intense at present, with both companies actively seeking opportunities to gain a strategic advantage over the other. A recent report indicated that OpenAI’s growth in Q2 was slower compared to Anthropic, whose annualized revenue run rate is now reportedly $65 billion. Anthropic investors have even suggested the company could achieve an IPO valuation of $2 trillion, while OpenAI is also reportedly working towards its own initial public offering.