[Into the World of AI] "No Access to Customer Data" vs. "30-Day Retention": Divergent Choices by OpenAI and Anthropic

OpenAI Testing Privacy Protection Processing Technology
Detecting Attacks Without Accessing Client Prompts or Responses
Anthropic Retains Client Data for 30 Days
Race for Enterprise AI Market Leadership Ahead of Public Listings

OpenAI and Anthropic have introduced different data management strategies in the enterprise artificial intelligence (AI) market. While both companies share the objective of detecting long-term cyberattacks executed by AI agents, they have adopted fundamentally opposite approaches to handling customer data.

OpenAI Pre-Released 'Privacy Protection Measures'. OpenAI Homepage.

According to industry sources on August 22, 2026, OpenAI is currently testing its 'Privacy Protection Processing' technology with select corporate clients, which automatically detects signs of cyberattacks and model abuse without viewing or storing users’ prompts and responses. The company plans to expand the scope of this technology in September and release a technical white paper.

The background for this development lies in the growing number of cyber risks that can only be detected by linking multiple conversations, as AI evolves into an agent capable of performing complex tasks. For instance, although individual messages may appear as ordinary questions from employees, patterns such as inquiries about software vulnerabilities, remote access, and methods to evade security detection, when found across multiple conversations, could indicate a long-term hacking attempt. Alea Haasz, Head of Product Policy at OpenAI, stated, "With our advanced models, risks often emerge not from any single prompt and response, but across multiple interactions observed over time."

This technology overcomes the limitations of the previous zero data retention (ZDR) policy, which made it difficult to detect risks spanning several conversations by not preserving customer data. Instead, the data remains on infrastructure controlled by the customer, while an automated system analyzes multiple interactions, transmitting only a limited set of signals, such as risk type and severity, without storing the actual prompts and responses. This approach enables the detection of prolonged attacks and malfunctions by AI agents while still adhering to a no-data-retention principle.

Anthropic: "Cyberattack Detection Requires 30-Day Data Retention"

Meanwhile, in June, Anthropic mandated a 30-day retention period for usage records of its top-tier models, Claude Phaible 5 and Mythos 5. The reason provided was that cyberattacks are difficult to detect based solely on individual queries, requiring records to be analyzed over a certain period. In a risk report published this month, Anthropic explained its data retention policy by stating that attacks leveraging stolen accounts or API keys against a company’s systems and data were only detectable because multiple requests made over time were aggregated.

However, corporate clients using Anthropic’s models have expressed concerns about storing confidential materials such as source code and customer information with an external service provider. It was reported that Microsoft raised issues with Anthropic’s policy, restricting some employees’ use of Phaible 5 and reviewing the potential legal implications.

On August 20 (local time), Anthropic announced its plan to maintain the 30-day retention policy but to modify it so that data can be kept in the client’s own cloud environment. The new approach would retain records necessary for security monitoring while granting clients control over the storage location and access rights.

Foreign media outlets have interpreted the two companies’ moves as a competition for dominance in the enterprise AI market ahead of their anticipated public listings. The Wall Street Journal (WSJ) evaluated OpenAI’s new technology as a move to attract enterprise customers dissatisfied with Anthropic’s data retention policy. In the market for enterprise AI, factors such as data management and security—beyond just model performance and cost—are rapidly emerging as new criteria in customers’ decision making.

There are also opinions cautioning against viewing data protection purely as an issue of storage or deletion. It is anticipated that going forward, competition will expand beyond simply whether to retain data to include the security of data processing and the ability to facilitate post-incident traceability. The Cloud Security Alliance (CSA) emphasized, "Encryption and access control measures should cover the entire data lifecycle, from collection and transmission to inference, storage, and disposal."

This content was produced with the assistance of AI translation services.