OpenAI Previews Private Safety Processing to Rival Anthropic
August 19, 2026
On August 19, OpenAI announced Private Safety Processing, a safety monitoring technology designed to detect misuse across interactions while remaining compatible with Zero Data Retention (ZDR).
The company is currently testing the system with early customers and plans to begin rolling it out in September.
The approach stands in contrast to Anthropic, which retains inputs and outputs of its most capable models for 30 days — a policy that has reportedly raised concerns among some enterprises.
How Private Safety Processing Works
OpenAI already offers ZDR to eligible API customers, using automated systems to evaluate each interaction individually for signs of abuse without storing customer data.
Private Safety Processing widens that scope. It assesses inputs and outputs across multiple conversations, not just one, to catch misuse that unfolds over time — for example, a bad actor spreading out malware-related requests to evade detection.
Under ZDR, customer content can stay on infrastructure the customer controls. OpenAI says it is also developing an option to store content on its own infrastructure, encrypted with keys the customer manages. OpenAI does not hold those keys, so in either setup its personnel cannot access the underlying content.
According to TechCrunch, a company spokesperson said that when the system is triggered, it may send a “narrowly defined signal” describing the type of activity. OpenAI can then contact the customer, and any data sharing remains at the customer’s discretion.
Anthropic’s 30-Day Retention for Its Most Capable Models
Anthropic retains prompts and outputs of what it calls covered models — its Mythos-class models and future models with similar capabilities — for 30 days. The policy took effect on June 9, 2026, to support the company’s safety work.
Anthropic says no personnel can read retained conversations by default. Human review occurs only through a controlled access path, is limited to a small set of approved reviewers, and every access is recorded in a tamper-proof log. The change mainly affects organizations that had been using Claude under zero-data-retention agreements. Consumer plans — Free, Pro, and Max — are unaffected because inputs and outputs are already retained there.
Still, TechCrunch reports that some enterprises handling sensitive data have raised significant concerns about the policy. Anthropic has been expanding other safety measures as well, including invisible watermarks for Claude-generated text.
Data Handling Becomes a Competitive Battleground
Long-running AI tasks can reveal patterns of risk that are difficult to detect when interactions are reviewed in isolation.
OpenAI says it is shaping Private Safety Processing with organizations that handle sensitive information such as financial records, health data, and confidential business plans. For companies facing regulatory obligations, whether an AI provider retains data can directly affect adoption decisions.
The two rivals are competing intensely for enterprise customers, and safety architecture itself is becoming a selling point.
What to Watch Next
OpenAI plans to begin rolling out Private Safety Processing in September, though it has not said which customers will qualify or how quickly access will expand.
Whether Anthropic adjusts its retention policy in response to enterprise feedback — and how the contrasting approaches influence enterprise procurement — remains open.
Sources
OpenAI – Offering Zero Data Retention for frontier models
Anthropic – Data retention practices for Covered Models
Bloomberg – OpenAI to Enhance Safety Processes for Paid Tool Customers
TechCrunch – OpenAI seeks to one-up Anthropic with new customer privacy protections
