Anthropic is preparing a safety system that would let enterprise customers keep data generated by its most advanced AI models on their own cloud infrastructure, while retaining the data for 30 days. The change would soften a policy introduced in June for traffic on the Fable and Mythos models and future frontier models, which Anthropic said was needed to detect sophisticated cyberattacks spanning multiple requests. More than 100 customers, including Salesforce, have helped develop the system, which is expected later this year. The policy has faced criticism from David Sacks, Satya Nadella and Alex Karp, who said enterprises need stronger assurances that providers will not store sensitive data. OpenAI has separately previewed Private Safety Processing, which is designed to identify misuse across interactions without retaining customer prompts or responses. The competing approaches are emerging as the companies prepare for potential public listings and vie for enterprise customers using AI with trade secrets, legal documents and proprietary research.