OpenAI and Anthropic Have a Zero Data Retention Dilemma—and Trap

For many enterprises processing sensitive customer data, the lack of zero data retention (ZDR) is a dealbreaker when it comes to deploying frontier AI models. That’s forcing AI companies to rethink their strategies. 

ZDR is a security practice in AI where providers do not store prompts or use them for model training, abuse monitoring, or product improvement beyond the immediate API call.

Anthropic had previously introduced a mandatory 30-day retention period for enterprise traffic on Fable 5 and Mythos 5, arguing that it needed the data to detect sophisticated misuse, including attacks and jailbreaks that could span multiple interactions.

It is now changing that approach.

According to a Bloomberg report, the company will continue requiring 30-day retention for its most advanced models, but will allow enterprise customers to keep the retained data on their own cloud infrastructure. This policy evolved after reportedly talking to more than 100 customers, including Salesforce.

OpenAI Announces Private Safety Processing

OpenAI is taking a different route.

Historically, its ZDR-compatible safety systems evaluated API interactions individually, processing data briefly in memory before discarding it. That protected customer data, but limited the ability to identify misuse that only becomes apparent across a series of connected interactions.

OpenAI is now testing “Private Safety Processing”, a framework designed to monitor multi-turn interactions while remaining ZDR-compliant. 

Customer prompts and completions remain on customer-controlled infrastructure or are protected using customer-managed encryption keys. Safety systems can analyse interaction streams and send only limited signals about potential misuse to OpenAI, rather than exposing the underlying content. 

OpenAI plans to begin rolling this out to enterprise customers in September.

The difference between the two approaches is important. Anthropic is moving the required retention closer to the customer. OpenAI is trying to make retention redundant.

Other cloud providers show that there is no single industry solution.

Microsoft allows eligible Azure customers to apply for Modified Abuse Monitoring, which can remove the standard abuse-monitoring retention and human-review process. But this is an approval-based exception, rather than a setting every customer can simply switch on. Microsoft currently limits the programme to customers and partners managed by a Microsoft account team or covered by an eligible programme.

Google takes a more granular approach. Its Vertex AI documentation allows customers to achieve ZDR for core model inference, but individual features can introduce separate retention requirements. 

Grounding (connecting with the AI model in real-time) with Google Search, for example, can retain derived queries and contextual information for up to three days, while Grounding with Google Maps retains prompts, context, and generated outputs for 30 days. Some advanced AI features may also have additional logging requirements.

AWS takes a different approach again. Amazon Bedrock operates with Zero Operator Access and ZDR by default, meaning AWS operators cannot access model inputs or outputs, and Bedrock does not normally store them. But model-specific safety requirements can override that baseline. 

However, Palantir is stricter, given its access to national security databases. Its AIP platform requires third-party AI models to meet its ZDR requirements regarding prompts and responses before it offers them to customers. 

Why Does This Matter

Data retention is becoming harder to enforce as models move from answering individual prompts to performing long-running, multi-step tasks. Moreover, it’s becoming harder to contain advanced AI agents in standard sandboxes because of dynamic code and unpredictable runtime tool calls, making any misuse harder to detect from individual interactions. 

When introducing its 30-day retention requirement in June this year, Anthropic explicitly cited that multi-session tracking would help identify complex threats or misuse that span multiple requests.

That creates a conflict, where model providers want more visibility into behaviour, while enterprises don’t want providers to access their data.

That may have been one factor limiting Fable 5’s enterprise adoption. The model accounted for just 11.4% of business spending on Anthropic models in its first two months, according to Ramp, despite being the company’s most capable model.

Cost was one factor, but its mandatory 30-day retention policy created another barrier for organisations that require ZDR.

Sensitive Data on a ‘Post-it Note’

“For highly regulated industries, using a model with a 30-day data retention policy is like leaving your most sensitive data on a post-it note in a public lobby and hoping for the best,” Sathya Narayanan Annamalai Geetha, Senior Principal Architect at Google and fellow of the British Computer Society, tells AIM.

He calls it an “automatic disqualifier” for such organisations.

“In the enterprise world, security and compliance decisions are binary,” he says. “If an enterprise requires ZDR to satisfy compliance, legal, or proprietary security mandates, a 30-day retention policy is a non-starter.”

The broader issue is that enterprise AI adoption is increasingly being decided by governance requirements rather than model capability alone.

“For many regulated and risk-sensitive organisations, data governance requirements can be a hard gating factor that outweighs both price and performance,” notes Biswajeet Mahapatra, Principal Analyst at research firm Forrester Research.

He says security, privacy, trust, and governance consistently rank among the biggest barriers to enterprise AI adoption, particularly at larger organisations with more extensive compliance obligations.

“Organisations increasingly apply AI governance frameworks, Zero Trust principles, data minimisation requirements, and approved tool policies before they assess the model's capabilities,” he says.

And this turns data retention from a technical specification into a procurement requirement.

“If a retention policy conflicts with an organisation's regulatory, contractual, or governance requirements, that model may effectively be excluded from consideration regardless of its performance or cost profile,” Mahapatra says.

Even as AI model providers address some of the practical barriers to enterprise adoption—from cutting API prices, as OpenAI did with an 80% reduction for GPT-5.6 Luna, to investing in forward-deployment efforts that help organisations integrate models into real-world workflows—data governance remains a harder problem to solve. 

Both model providers and enterprises will need to find terms that balance the providers’ need to monitor and govern their models with enterprises’ requirements around data privacy, security, and control.