OpenAI launches private safety checks that watch for abuse without storing your data

Key points

• OpenAI is previewing Private Safety Processing, a new automated safety system that monitors for misuse without retaining any customer data.
• Anthropic announced a policy in July 2026 to hold user conversation data for 30 days on what it calls "covered models," a move that has unsettled some enterprise customers.
• Anthropic's annualised revenue run rate is reportedly $65 billion, ahead of OpenAI's growth pace in Q2 2025.
• If OpenAI's system flags suspicious behaviour, it sends a narrow warning signal to the company, but no human reads your conversations unless you choose to share them.
• Both companies are heading toward IPOs, making the race for enterprise trust a financial priority, not just a technical one.

OpenAI wants enterprise customers to know this: it can watch for bad behaviour without ever reading what those customers type.

The company is previewing a new service called Private Safety Processing to a select group of customers. Think of it as a security camera that records suspicious movement but wipes the footage every session automatically, unless it spots something genuinely alarming.

What does this actually mean for businesses using AI?

For any company worried about sensitive data landing in an AI lab's hands, this matters. Private Safety Processing extends what is already standard practice at OpenAI.

That existing practice is called Zero Data Retention, or ZDR, a policy where AI software agents (automated programs, not human staff) scan each conversation for abuse in real time and then discard it. No storage, no human eyes. We first covered ZDR on 19 August 2026.

Private Safety Processing goes further. It watches patterns across multiple conversations over time, not just a single session, using an automated agent to look for signs that someone is slowly building toward something harmful. Spreading requests across sessions to avoid detection is a known tactic: someone quietly asking an AI to help design malicious software one step at a time, for instance. This system is designed to catch that.

If the agent spots a problem, it sends a "narrowly defined signal" to OpenAI, flagging the type of concern without forwarding the actual conversation. OpenAI then decides whether to act, and if it does, it reaches out to the customer for context. You choose whether to hand over any data at that point.

Why is OpenAI doing this now?

It's a direct response to Anthropic, its closest competitor. As first reported by TechCrunch AI, the timing lines up neatly with customer friction over Anthropic's newer data policy.

In July 2026, Anthropic announced it would retain customer conversation data for 30 days on what it calls "covered models," which include its most capable AI systems. A small team of approved reviewers can access that data through a controlled process, with every review logged in a tamper-proof record. Anthropic's stated reason is safety: having those conversations lets it investigate potential abuse.

Some enterprise customers in regulated fields handle information so sensitive that even a 30-day retention window at a third-party AI lab is a hard no.

The business stakes are real. Anthropic's annualised revenue run rate sits at roughly $65 billion, and investors have floated a $2 trillion valuation ahead of a possible public listing. OpenAI is also working toward an IPO, and it's doing so through significant leadership change: our recent reporting tracked two senior executives out in a single week as that process accelerates. Winning privacy-conscious enterprise clients could move those numbers meaningfully.

Should businesses act on this right now?

If your company uses an AI API (a technical connection that lets your product talk to an AI model) to process sensitive data, put two direct questions to your provider: do you retain my data, and under what conditions can a human at your company read it? Don't wait for a sales call to raise them.

OpenAI's Private Safety Processing is still in preview, not yet available to everyone. But the direction is clear: privacy is becoming a competitive selling point, not just a compliance checkbox.

Here's my honest read: the companies that will win enterprise AI contracts over the next two years aren't necessarily the ones with the best models. They're the ones that can credibly prove they won't mishandle what flows through those models. That's a harder thing to demonstrate than a benchmark score, and the press release is always the easiest part of it.