- OpenAI has introduced the Private Safety Processing system, which enables the detection of abusive behavior on advanced AI models without storing customer data.
- The technology is currently being tested with selected enterprise customers before a wide release alongside technical documentation in September.
- The system only sends limited safety signals to OpenAI instead of transmitting entire user prompts or responses.
- Data can remain on customer-controlled infrastructure or be stored by OpenAI using encryption keys managed by the customers themselves.
- OpenAI stated that tracking multiple interaction sessions over time helps detect complex attacks, such as users collecting fragments of information to prepare for a cyberattack.
- In contrast, Anthropic requires enterprise customers using Fable 5 and Mythos 5 models to accept data retention for 30 days for safety monitoring purposes.
- Anthropic admitted that this policy might not be supported by customers and could create a competitive disadvantage if rivals maintain no-data-retention models.
- The two companies are pursuing different safety strategies as AI models become more powerful and new risks emerge.
- OpenAI also recently paused part of the training process for new models to address safety issues, while Anthropic stated it sees no need for similar measures.
- OpenAI’s zero-data-retention system is currently available only for enterprise customers and eligible APIs, not for ChatGPT Free, Go, Plus, or Pro users.
📌 The competition between OpenAI and Anthropic is expanding into how they balance privacy and AI safety. OpenAI aims for risk detection while maintaining a no-data-retention mechanism for enterprise customers, whereas Anthropic accepts 30-day logging to track sophisticated attacks. These two approaches reflect different choices in AI governance as models become increasingly capable and pose higher risks.
