OpenAI Found a Way to Watch You Without Looking

Thirty days. That is how long Anthropic can now hold onto an enterprise customer’s sessions and the conversations inside them when those sessions run on a “covered model.” The policy went live in July, and it did exactly what a 30-day retention window does to a bank, a hospital, or a law firm: it made them nervous. On August 19, OpenAI walked into that nervousness with a product.

What OpenAI Actually Shipped

Zero Data Retention, stretched across sessions

Zero Data Retention is not new. Under ZDR, an eligible API customer gets a simple promise: OpenAI does not keep prompts or model responses after a request is processed, OpenAI personnel cannot review that content, and enterprise data is not used to train models unless the customer explicitly opts in. Abuse monitoring still happens, but an automated agent handles it inside a single session and then the data is gone.

The gap in that design is obvious once you think like an attacker. If detection only ever looks at one conversation at a time, an adversary simply spreads the work out. Ask for one fragment of malware logic here, another there, across a dozen separate sessions, and each individual request looks unremarkable. OpenAI’s spokesperson described exactly this scenario to TechCrunch: a bad actor spreading requests out to avoid detection.

Long-horizon monitoring without a paper trail

Private Safety Processing is OpenAI’s attempt to close that gap without giving up the retention promise. The company describes it as long-horizon safety monitoring that assesses inputs and outputs across multiple related conversations rather than one. An agent does the correlating. If something trips, OpenAI receives what it calls a “narrowly defined signal” warning of a specific type of activity — not the prompts, not the responses.

From that signal, OpenAI decides whether enforcement is necessary. If it is, the company contacts the customer for context or to work through the issue together, and the customer may choose to share data at their own discretion. Customer data itself can stay on customer-controlled infrastructure, or sit with OpenAI under encryption keys the customer holds.

Why Anthropic’s 30 Days Became a Liability

The covered-models carve-out

Anthropic largely abides by Zero Data Retention too. The exception is the category it calls “covered models” — all Mythos-class models and, in the company’s own wording, “future models with similar capabilities.” For those, sessions and the conversations within them are retained for 30 days so the lab can sift for potential impropriety. The rationale is safety at the frontier: the more capable the model, the higher the cost of missing an abuse pattern.

The rationale did not land evenly. The Wall Street Journal reported a brewing backlash in Silicon Valley, and the specific objection is not hard to reconstruct. An enterprise that processes regulated customer records does not want those records sitting in a third-party lab for a month, and it particularly does not want them inspected. Anthropic’s counter is procedural rather than technical: human review can happen, but only through a controlled access path involving a small set of approved reviewers, with every review session recorded in a tamper-proof log that reviewers cannot suppress or modify.

Two philosophies, one procurement checklist

Strip away the marketing and this is a genuine disagreement about where trust should live. Anthropic’s answer is auditable human process — someone may look, and there will be an unforgeable record of it. OpenAI’s answer is cryptographic and architectural — nobody looks, because the system is built so that only a flag escapes. Neither is obviously correct. A tamper-proof log gives you recourse after the fact; a design that never surfaces content gives you nothing to recover from because nothing was exposed.

The Competitive Subtext

A revenue gap that explains the timing

This announcement did not arrive in a vacuum. Anthropic’s annualized revenue run rate is now reported at $65 billion, and a Wall Street Journal report showed OpenAI’s second-quarter sales growing more slowly than Anthropic’s. Anthropic investors have floated an IPO at a $2 trillion valuation; OpenAI is working on its own IPO, with CNBC reporting timing discussions pointing toward 2027. When two companies are both walking toward public markets and one is growing faster, every differentiator gets weaponized — including the boring ones.

Privacy is an unusually effective weapon in enterprise sales precisely because it is boring. It does not require the buyer to believe your model is smarter. It requires only that their general counsel prefers your paperwork. That is a much lower bar to clear, and it is durable in a way that benchmark leads are not.

What is still unproven

Private Safety Processing is a preview, running with early customers, and it is scoped to eligible enterprise and API accounts — not consumer ChatGPT subscribers on paid plans. The technical white paper is scheduled for September, and until it lands, the strongest claims remain unverified from the outside. Cross-session correlation that never exposes content is a hard engineering problem, and the specific question buyers should ask is how coarse that “narrowly defined signal” really is. A signal precise enough to be actionable is, by definition, a signal that carries information about the underlying content.

There is also an enforcement asymmetry worth noting. Under OpenAI’s model, when a signal fires, the company has to ask the customer for context, and the customer may decline. That is excellent for privacy and awkward for safety. Anthropic’s retention window exists precisely so the lab does not have to ask permission to investigate. Whether regulators eventually prefer the version where the AI lab can look, or the version where it structurally cannot, is an open question that neither company controls.

Related

• Anthropic Hid Something in Every Word Claude Writes
• GitHub Went Dark. Cursor Was Already Waiting.
• 60 AI Agents Just Attacked Math’s 150-Year Mystery
• Meta Just Made the Cloud Optional for AI Agents
• Claude Code Will Stop Asking Your Permission Soon

Sources

1. TechCrunch — OpenAI seeks to one-up Anthropic with new customer privacy protections (Aug 19, 2026)
2. OpenAI — Offering Zero Data Retention for frontier models
3. Axios — OpenAI previews zero-retention safety system as Anthropic requires data logs
4. Anthropic — Data retention practices for covered models