Your Enterprise Data Just Became the Real Battleground Between OpenAI and Anthropic
OpenAI unveiled Private Safety Processing Wednesday, a system for detecting AI misuse without retaining customer data, directly challenging Anthropic's 30-day data retention rule.
OpenAI announced Wednesday that it is previewing Private Safety Processing, an enterprise safety system designed to detect misuse across multiple conversations without storing the underlying prompts or responses.
It directly challenges the approach of its rival Anthropic for its most capable models. The announcement comes the same month OpenAI disclosed it was pausing training on parts of Astra over emerging cyber risks.
It also highlights a growing divide between the two labs: whether AI safety requires retaining customer data or can work without seeing it at all.
Watching Behavior Without Ever Reading the Conversation
OpenAI’s Head of Product Policy Aleah Houze said risks can emerge from patterns across multiple interactions rather than a single prompt, Axios notes.
She cited a scenario in which someone might ask about a software vulnerability in one conversation, as AI is now quick at this, and then return later asking about remote access tools in another.
Private Safety Processing is built to catch exactly that pattern, TechCrunch reported. An automated agent that flags cross-session activity and sends OpenAI a limited safety signal rather than the actual conversation content.
Customers can keep their data on their own infrastructure or let OpenAI store it using encryption keys they control.
OpenAI says a full technical white paper and broader rollout are coming in September.
Two Labs, Two Answers to the Same Threat
Both companies are tackling the same problem: sophisticated attacks that unfold across separate harmless-looking requests, but they disagree on how to catch them.
Anthropic’s risk report last week defended its 30-day retention policy for Fable 5 and Mythos 5, acknowledging that it may upset customers used to zero retention and hurt the business if rivals do not follow.
Anthropic argues the policy is still needed to detect attacks spanning multiple requests.
OpenAI’s Wednesday announcement reads as a direct response: it is betting the same threats can be detected without asking enterprise customers to accept the retention risk Anthropic considers necessary.
Why This Fight Is About Trust as Much as Technology
The stakes go beyond which system catches more misuse.
Enterprise customers that use AI in legal, healthcare, and financial sectors have faced growing demands for broader data access to serve them better, while Anthropic’s retention policy has already drawn backlash from some customers.
If OpenAI’s zero retention approach works as described, data privacy could become a competitive advantage as both companies pursue IPOs and the same enterprise contracts.
Anthropic’s annualized revenue run rate has reportedly reached $65 billion, while OpenAI’s Q2 growth trailed its rival.
The bigger question is whether Private Safety Processing can detect sophisticated attacks without full conversation context. Its privacy benefits are plausible, but the real test is whether limited safety signals can catch threats as reliably as Anthropic’s broader data access.
