OpenAI seeks to one-up Anthropic with new customer privacy protections

OpenAI is taking a swing at Anthropic with a new privacy pitch for enterprise customers: watch for abuse, but don’t keep the data. The company is previewing Private Safety Processing to select customers, and says the system can flag possible misuse while retaining none of a customer’s information. That’s a clean contrast with Anthropic’s newer retention rules, which have already annoyed some big customers.

The timing matters because the pressure around AI safety is rising as models get more capable. Companies want guardrails, but enterprise buyers also want their sensitive material treated like sensitive material, not stored for later inspection. Anthropic’s July policy allows it to keep user data for 30 days for “covered models,” including Mythos-class models and “future models with similar capabilities.” The company says that helps it look for abuse. Customers, especially those handling confidential data, heard something closer to surveillance.

OpenAI says it already uses Zero Data Retention, or ZDR, for a relative level of privacy. In that setup, agents inside the OpenAI API monitor a single session for abuse without human review and without the company holding onto customer data. Private Safety Processing is meant to push that further. OpenAI describes it as long-horizon monitoring that can examine inputs and outputs across multiple conversations, not just one.

That distinction is the whole point. A bad actor trying to build malware or carry out a cyberattack might spread requests over several sessions to stay under the radar. OpenAI says the new system can catch that pattern without a person reading through the chats. If it triggers, the company may receive only a narrowly defined signal about a specific kind of activity, then decide whether enforcement is needed.

Anthropic still allows human review in some cases, but says it happens through a controlled access path and is logged in a tamper-proof way. Behind all the privacy language is a tense rivalry. A recent report said OpenAI’s Q2 growth was slower than Anthropic’s, Anthropic’s annualized revenue run rate is now reportedly $65 billion, and Anthropic investors have floated a $2 trillion IPO. OpenAI is working on one too. Nobody likes a customer revolt, but everyone loves a benchmark, a revenue number, and a chance to say the other side is being reckless.