On August 21, Bloomberg reported that Anthropic is preparing to allow enterprise customers to store their data on their own cloud infrastructure instead of Anthropic's servers—a major reversal from its June policy requiring 30-day retention for all Mythos and Fable series model interactions. The new security system is expected to launch later this year and is being developed with over 100 customers in highly regulated industries.
The shift comes just weeks after OpenAI announced Private Safety Processing, a system that monitors for AI misuse across multiple conversations without retaining customer data. Microsoft and Databricks are already testing it. Anthropic's reversal confirms that a competitor's product strategy has changed the baseline of what enterprise customers expect.
OpenAI Changed the Game. Anthropic Is Playing Catch-up.
In June, Anthropic justified its data retention policy as a safety necessity. The company argued that sophisticated misuse patterns—like state-sponsored espionage—only become visible across multiple requests over time, making data retention "not optional." It warned the policy "will be unpopular with customers who have come to expect zero retention" and "pose real risks to our business success (especially if competitors do not follow)."
But when OpenAI launched Private Safety Processing in August, it presented a direct challenge to that assumption. OpenAI's system analyzes cross-session patterns for risk signals without retaining the underlying data. The pitch: safety monitoring doesn't require data ownership. Anthropic's reversal is a direct admission that its competitors have validated a different approach—and that customers are choosing that approach.
Safety Doesn't Require Owning Your Data
The new system represents a fundamental shift in Anthropic's thinking. The company previously argued that to detect cross-session network attacks, you had to string together dozens or even hundreds of requests—meaning the data had to be stored and accessible on Anthropic's side.
But the new system decouples safety from data ownership. Customers keep their data in their own infrastructure while Anthropic's automated systems still analyze patterns and flag risks. It's the same safety outcome without the data transfer. The architecture suggests that Anthropic has found a way to do the cross-session analysis without physically holding the data—something it had previously claimed was impossible.
The detail that matters most: Anthropic began developing this system months ago, before the customer backlash intensified. The company had already identified the flaw in its own policy. The market reaction just accelerated the timeline.
Enterprise Customers Are Voting With Their Budgets
Anthropic's reversal is not just about policy. It's about market dynamics. The company's Q2 revenue of $116 billion briefly surpassed OpenAI's $67 billion, but quarterly data from Ramp shows that OpenAI's enterprise customer growth has now overtaken Anthropic's—the first time in 2026 that OpenAI has led in new customer acquisition.
The trend lines tell a clear story: customers in regulated industries—healthcare, finance, government—are choosing vendors based on data sovereignty as much as model capability. Microsoft CEO Satya Nadella and Palantir CEO Alex Karp have both raised concerns about data retention policies, arguing that highly regulated industries require clear "no data storage" commitments.
Anthropic's policy reversal is a concession to that market reality. The company needs to protect its enterprise growth while maintaining its safety narrative. But the dual-track strategy introduces new questions: will customers trust the new system as much as a true zero-retention policy? And will Anthropic now have to prove that its safety monitoring works without the data it originally claimed was necessary?
P.S. The system is still in development. Anthropic plans to roll it out to enterprise customers in the coming months, with a focus on regulated industries. But the announcement itself—and the speed with which it followed OpenAI's launch—says more than the technical details. The AI safety debate has moved from "who can build the most powerful model" to "who can protect the most sensitive data." And in that debate, Anthropic is playing catch-up.
Frequently Asked Questions
Q: What is Anthropic's data retention policy changing?
A: Anthropic is developing a new security system that will allow enterprise customers to store 30 days of their AI data on their own cloud infrastructure, rather than on Anthropic's servers. The system has been in development for months and is expected to launch later this year.
Q: Why did Anthropic implement the 30-day retention policy in June?
A: Anthropic implemented the policy for its most capable models—Mythos and Fable series—to detect sophisticated misuse patterns that only become visible across multiple requests over time. The company argued that threats like state-sponsored espionage require analyzing conversations together over a 30-day window.
Q: What triggered Anthropic's reversal?
A: Two factors: customer backlash (Microsoft CEO Satya Nadella and Palantir CEO Alex Karp both raised concerns), and OpenAI's August 19 announcement of Private Safety Processing—a system that monitors for AI misuse across conversations without retaining customer data. OpenAI's approach set a new baseline for enterprise AI procurement.
Q: How is Anthropic's new system different?
A: It keeps the 30-day safety monitoring window but moves data storage to customer-controlled cloud infrastructure. Customers keep their data; Anthropic still performs cross-session analysis for safety. The architecture decouples safety monitoring from data ownership.
Q: Was this system developed in response to the backlash?
A: No. Anthropic began developing the system months ago, before much of the backlash intensified. The company had already identified the flaw in its own policy; the market reaction accelerated the timeline.
Q: What does this mean for enterprise customers?
A: Enterprise customers in regulated industries—healthcare, finance, government—will soon be able to use Anthropic's most powerful models while storing data in their own infrastructure. This addresses data sovereignty concerns while maintaining safety monitoring.
Q: How does Anthropic's new system compare to OpenAI's Private Safety Processing?
A: OpenAI's system operates on a true zero-retention basis—it does not store customer data at all. Anthropic's system still requires a 30-day retention window but allows customers to store that data in their own cloud. The difference is where the data physically resides, not whether it's retained.
Q: What is the "zero retention" standard?
A: Zero retention means the AI provider does not store customer prompts or model responses after processing. OpenAI's Private Safety Processing operates on this model. It has become an increasingly important requirement for enterprise AI procurement, especially in regulated industries.
Q: How are customers responding to Anthropic's change?
A: Customer reception remains mixed. Regulated industries that need to control their own data and comply with privacy laws are relieved by the flexibility to retain data in their own cloud. However, some customers still prefer true zero-retention approaches like OpenAI's, rather than even a 30-day retention window, regardless of where the data is stored.
Q: What happens next?
A: Anthropic plans to roll out the system in the coming months, starting with enterprise customers in regulated industries. The company is also considering further adjustments to meet customer expectations. But the direction is clear: Anthropic is accepting that safety monitoring and data ownership are no longer tied together.
