Anthropic announced that, following a wave of feedback from enterprise customers, it is changing its controversial data retention rules and replacing them with a solution called Enterprise Frontier Safeguards. Under the new arrangement, companies will be able to retain data in their own cloud infrastructure, such as Amazon S3, Azure Blob Storage, or Google Cloud Storage, using their own encryption keys and access policies. Automated security checks will continue to look for signs of abuse, but no Anthropic employee will have access to the data.
Why the data was retained
The original rules took effect this June. Anthropic introduced them alongside the launch of the Claude Fable 5 and Mythos 5 models and required all traffic involving these models to be retained for thirty days. It justified this as a defense against abuse and new, previously unknown cyberattacks.
The company argued that some attacks can only be identified by examining a larger number of requests. As an example, it cited so-called best-of-N safeguard bypassing, in which an attacker submits hundreds of minor variations of the same prompt and hopes that one gets through. Larger campaigns, such as state-sponsored espionage or extortion using stolen data, only become apparent when classifiers can compare multiple requests with one another. According to Anthropic, the most sophisticated abuse spans dozens of tasks, different sessions, and accounts, so analyzing individual interactions followed by immediate deletion is not enough. Detection simply needs a window of time to connect the dots. The company also reiterated that its goal was not to obtain training material, adding that it has never trained and will never train on enterprise data without explicit consent.
Customers, however, were mainly concerned that their prompts and responses would be stored somewhere for a month. The companies Anthropic spoke with generally understood the security benefits of retaining data. Many of them, particularly those bound by strict industry regulations, nevertheless could not afford to use such models. Kate Jensen, who heads Anthropic’s U.S. commercial division, said that finding a different solution required hundreds of hours of work with customers. According to her, the company can now perform the necessary checks directly within customers’ systems and data, so no one has to compromise their privacy principles.
So what is changing
Enterprise Frontier Safeguards stores monitoring data in a cloud controlled by the customer, not Anthropic. Detection remains Anthropic’s responsibility, but data management, keys, and decisions about access remain with the company using the service. Anthropic does not charge for this service, and the checks work the same way whether the customer uses the technology directly or through a cloud platform. The storage itself, however, is billed by the regular cloud provider, including capacity, reads, writes, and outbound data transfers, just like any other item. Deployment is taking place in phases, with broader availability planned for this fall, while customers whose turn has not yet come are guaranteed zero data retention for the Fable 5 and Fable 5.1 models.
The original June rules remain in effect for customers outside enterprise programs who use Mythos-class models. The standard Claude Free, Pro, and Max plans are not affected by the change because inputs and outputs under these plans were already being retained.
The documentation for the original rules also describes safeguards intended to alleviate concerns. By default, no one at Anthropic can access stored conversations. A person can open them only through an approved process, typically when automated systems flag the content as potentially harmful, and access is restricted to a small group of vetted employees. Every such access is recorded in a log that cannot be deleted or modified. After thirty days, the data is automatically deleted, except in reported cases or situations where the company is legally required to retain it.
New models have been released
Alongside the announcement, Anthropic also introduced the Claude Fable 5.1 and Claude Mythos 5.1 models, which it describes as its best tools for programming, knowledge work, and research. Both are built on the same foundation and differ in the strictness of their safeguards. Fable 5.1 is generally available, while Mythos 5.1 operates only in trusted access mode and offers protections focused on cybersecurity and the biological sciences. Anthropic CEO Dario Amodei outlined the new rules last week when discussing an expanded partnership with Salesforce, emphasizing the effort to keep customer data private and models under control. Enterprise contracts are Anthropic’s main source of revenue, and the company is also targeting this market ahead of its planned stock market debut. Additional pressure is coming from competitors, as OpenAI introduced its own data retention solution earlier this month.
Sources: cnbc.com and finance.yahoo.com
