OpenAI Launches 'Private Safety Processing' System to Protect Customer Privacy

As the capabilities of artificial intelligence models grow, so too have the risks of their misuse, which has intensified calls for safety controls capable of preventing such abuses. Here, AI companies find themselves facing a delicate equation: how to balance respecting their corporate customers' privacy with monitoring usage to detect any potential violations?

In a move aimed at strengthening its competitive position against its rival Anthropic, OpenAI has announced a new approach to monitoring misuse that places privacy at the forefront of its priorities. The company is launching an early preview of a service called "Private Safety Processing" for a select group of customers—an automated system that detects potential violations without retaining any part of the customer's data.

The Privacy Versus Safety Equation

This system stands in stark contrast to the data retention policy that Anthropic recently announced, which drew the displeasure of some of its customers. Under that policy, the lab is entitled to retain users' data—including their entire sessions and the conversations within them—for thirty days when it comes to what are called "covered models," a category that includes the most capable models and similar future models.

Although this policy, announced in July, was designed for safety purposes to allow the analysis of suspicious uses, it raised deep concern among some organizations that handle massive amounts of sensitive data and do not wish to have it stored or examined by an external party.

Beyond the 'Zero Data Retention' Policy

In fact, OpenAI—like most companies in the sector—grants its customers a level of privacy through a policy known as "Zero Data Retention." This policy relies on software agents within the API to detect misuse on a session-by-session basis, so that the company does not retain customer data while remaining able to examine suspicious activity without human intervention. Anthropic largely adheres to this policy as well, except when it comes to "covered models."

OpenAI describes its new service as an extension of the zero data retention policy, as it represents a form of long-term security monitoring that evaluates the inputs and outputs of multiple conversations, not just a single conversation. This monitoring remains in the hands of a software agent that captures interactions when triggered and analyzes them across multiple sessions in search of any indicators of misuse.

Detecting Threats Distributed Across Multiple Sessions

According to a company spokesperson, the new technology helps detect malicious uses that span multiple sessions. An attacker—such as someone trying to develop malware for a cyberattack—might spread their requests across several conversations to evade detection. Here, the Private Safety Processing system enables the analysis of those scattered conversations in search of evidence of abuse, without human review of the content of the user's conversations.

If the system is triggered, it may send a "narrowly defined signal" to OpenAI alerting it to a specific type of activity, based on which the company decides whether taking action is necessary. If so, OpenAI reaches out to the customer to obtain additional context or to collaborate with them in addressing the matter, with the decision to share data remaining in the hands of the customer.

By contrast, Anthropic notes that human review of customer data may occur, but through a "controlled access path" limited to "a small group of authorized reviewers," with every review session recorded in a tamper-proof log that reviewers cannot delete or modify.

Fierce Competition for the Lead

This move reflects the intense competition between OpenAI and Anthropic, as each seeks to seize any opportunity to gain an edge over the other. Reports indicate that OpenAI's growth rates in the second quarter were slower than those of its counterpart Anthropic, whose annual revenue rate is said to have reached around $65 billion.

• Anthropic may go public with a valuation of $2 trillion, according to its investors.
• OpenAI is also working on an anticipated initial public offering.
• Privacy has become a key competitive card in the enterprise AI market.

Ultimately, it appears that the battle to win the trust of major companies will not be decided by technical capabilities alone, but by the extent to which each provider can offer solid guarantees to protect its customers' most sensitive data.

✦ بقلم فريق دروب أيديا