OpenAI unveils "Private Intelligence" for sensitive work; new feature protecting data during processing in API preview

OpenAI unveils "Private Intelligence"

OpenAI on Oct. 30 (Oct. 31 Japan time) unveiled a set of initiatives that let companies use its top AI models for highly sensitive work as "OpenAI Private Intelligence," on its official X account and a business page. Its pillars are "zero data retention," which leaves no prompts (instructions to the AI) or responses behind after processing for covered work, and "Private Inference," which runs AI in an environment where data cannot be seen from outside even during processing. Private Inference is in preview in the developer API, and the company is accepting applications from interested businesses. OpenAI said business data, including inputs and outputs, belongs to the customer, and that customers can also decide which data AI agents (features in which AI carries out work on its own) can touch, what they can execute and when human approval is required.

OpenAI had given the initiative its name at its developer conference "DevDay 2026" on Sept. 29 (Sept. 30 Japan time) and had said it would offer a preview of Private Inference "this fall." This time, it set up the business page and an application point and announced it again. The business page says it will let companies use the most capable models for work they would not normally send to another company's API.

The text of the post could not be opened directly from this environment and was confirmed in a RuntimeWire article quoting it.

New feature protecting data during processing, details not disclosed

The new pillar, Private Inference, uses a technology called "confidential computing." It creates an area within a chip whose contents cannot be seen from outside and performs AI computation inside it, protecting data with encryption not only when it is stored or transmitted but also while it is being processed. OpenAI said it has strict, verifiable controls.

However, it has not disclosed which models and APIs are supported, whether the business-oriented "ChatGPT Enterprise" is covered, which chips are used, or what customers can verify for themselves. It has also not disclosed pricing, and the business page directs inquiries to its sales team. VentureBeat noted that such details determine the actual level of protection.

Safety monitoring under zero retention, with a copy kept on the customer side for 30 days

The other pillar, zero data retention, is a commitment that OpenAI has long offered to eligible business and API customers. Prompts and responses are not kept after processing, and employees do not see their contents. Unless the customer opts in, they are not used for model training either. However, content suspected of being child sexual abuse material is kept even under zero retention for reporting required by law.

On Aug. 19 (U.S. time), OpenAI had announced a preview of "Private Safety Processing," a system for detecting misuse without keeping data. In addition to individual exchanges, it automatically examines multiple related exchanges together, and if there is a problem, it tells OpenAI only the type and severity of the alert. Employees do not see the original content even when an alert is raised.

According to VentureBeat, content selected for safety review is encrypted and placed in the customer's cloud storage (AWS, Microsoft Azure, Google Cloud), where it is deleted after 30 days. It can be decrypted only by a review environment whose integrity has been attested by hardware, the publication said. As a result, even under "zero retention," an encrypted copy remains on the customer side for 30 days, and customers must keep managing the storage and keys during that time. The publication noted that this could affect internal and regulatory data retention policies.

Customers set agent permissions, but critics say technology alone cannot protect data

The business page said that for AI agents, customers decide which data and functions they may touch, which operations are permitted and when employee approval is required. It also cited data encryption, strict access controls and continuous threat detection. For third-party audits and certifications, it listed SOC 2 Type 2, ISO, HIPAA, the U.S. rules for handling health information, and FedRAMP, the U.S. government's cloud certification.

Some point to the limits of confidential computing. Matthew Green, who researches cryptography at Johns Hopkins University, said that once agents can read private data and send messages externally, technology that protects data during processing alone provides "effectively no" protection. German IT outlet All-AI.de said that even with the system in place, customers ultimately have to trust OpenAI.

Eligibility: API customers covered, consumer plans excluded

Zero data retention and Private Safety Processing are available to API customers who meet the conditions. Private Inference is a preview by application, and no timing for general availability has been given. According to Scalevise, the initiative as a whole will not launch all at once either, but will expand gradually from early adopters. The business page does not say whether Private Intelligence can be used in Japan. According to Axios, zero retention covers business and API customers, and settings for individuals using ChatGPT Free, Plus, Go and Pro do not change.

Anthropic requires 30-day retention for its top models

OpenAI and Anthropic have taken different approaches to handling corporate data. Anthropic requires even enterprise customers using its top models, "Fable 5" and "Mythos 5," to have prompts and responses retained for 30 days. According to The Register, the policy took effect on June 9 (U.S. time). Anthropic said it was necessary to detect sophisticated attacks split across multiple sessions and acknowledged that it will be unpopular with customers used to zero retention. The policy drew criticism from David Sacks, who oversees AI policy at the White House, Microsoft CEO Satya Nadella and Palantir CEO Alex Karp.

Large customers also took action. According to Bloomberg Government, relaying a report by The Information, NVIDIA, Palantir and Booz Allen Hamilton have sought new guarantees and reduced or stopped their use of OpenAI's and Anthropic's frontier models. Palantir reportedly obtained an "irrevocable" zero-retention guarantee from Anthropic for all models. OpenAI's announcement followed amid these spreading concerns.

Anthropic gains share of enterprise spending

The enterprise business has become a pillar of OpenAI's growth. According to Decrypt, the company's chief revenue officer, Denise Dresser, said in April that enterprise accounted for more than 40% of revenue and was expected to match consumer revenue by the end of 2026. Paid business users rose from 5 million in August 2025 to 9 million in February 2026.

In corporate spending, however, Anthropic has been gaining share. According to an index compiled by corporate card company Ramp based on transactions on its cards, OpenAI's share of spending on AI models fell from 44.2% in late September to 35.1% in late October. Anthropic's share rose from 51.3% to 60.5%. Whether the no-retention policy changes this trend will depend on how widely it is adopted once Private Inference's pricing and supported models are settled.

Ramp's figures count only payments by companies using Ramp's corporate cards.