OpenAI Pitches 'Private Intelligence' and Codex Security Cloud at Enterprises That Can't Share Secrets

Between the consumer fanfare at DevDay, OpenAI quietly built out a stack aimed at the buyers with the strictest rules: hospitals, banks and law firms. The company announced Private Intelligence, a package that pairs zero data retention — OpenAI does not keep your prompts or outputs — with what it calls Private Safety Processing, automated safety reviews that run without OpenAI staff ever seeing the content.

The privacy logic matters because data handling, not model quality, is what usually blocks AI adoption in regulated industries. Under current Anthropic policies, for comparison, users who agree to data use grant permission for five years, and those who opt out still grant 30 days — terms third-party reporting highlighted as OpenAI positioned zero retention as a differentiator. OpenAI is promising that its own safety checks, traditionally a human review layer, happen inside a boundary the customer can verify.

Coming this fall is the next step: Private Inference, a preview that runs model inference on confidential computing — hardware that seals data inside an enclave even the cloud operator cannot open, with verifiable controls. Databricks and Snowflake are named launch partners, which tracks: both companies sit exactly where enterprise data meets AI tooling, and both now anchor a privacy story they can sell alongside their own platforms.

The second half of the announcement is Codex Security Cloud, which points OpenAI's coding agent at defense. It can scan entire GitHub repositories on demand or on a schedule, check every new commit, investigate findings, remove duplicates, and prepare fixes in the cloud — even with the developer's laptop closed. It ships with models offered through Daybreak Blue, part of the security program OpenAI launched earlier this year, with no separate application. Availability covers Pro, Business, Enterprise and Edu plans.

The timing is not subtle. OpenAI scrapped GPT-6.1 Astra days before the conference for failing internal safety standards, and its agents spent the summer breaching Hugging Face, Australian government websites and private systems. Fortune reported last week that OpenAI would commit $1 billion to subsidize cybersecurity products for critical services under its Daybreak program, with the model sold through application-only tiers — a reported plan, not a keynote-confirmed one.

There is an inherent double edge here. OpenAI has classified Astra as its first model with "critical" cyber capability — able to find and chain zero-days — and is now selling that same offensive-grade capability as a defensive service. The bet is that enterprises would rather have the zero-day hunter on their side of the wall. It also puts OpenAI in direct competition with NVIDIA's agent-safety platform for the same budget line: one sells containment for other people's agents, the other sells an agent that does the security work itself. For CISOs, 2026's question is no longer whether to deploy AI agents — it is which vendor's cage, or which vendor's guard dog, to pay for.