OpenAI debuts Private Safety Processing to counter Anthropic's data retention policy
OpenAI is introducing a new feature focused on privacy, known as Private Safety Processing, to a select group of customers. This automated tool is designed to detect AI misuse without keeping customer information, analyzing inputs and outputs over various interactions to identify slow-developing malicious behavior. This initiative poses a challenge to Anthropic, which has been criticized for its data retention policy from July, mandating the storage of user sessions for 30 days for certain models. Although Anthropic adheres to Zero Data Retention (ZDR) principles, it makes exceptions for models like Fable. OpenAI expands its ZDR approach to include long-term monitoring, signaling potential concerns. The rivalry escalates as Anthropic announces a $65 billion revenue run rate and a possible $2 trillion IPO valuation, while OpenAI gears up for its own IPO.
Key facts
- OpenAI is previewing Private Safety Processing, a privacy-centric safety service for select customers.
- Private Safety Processing monitors for misuse without retaining customer data.
- The system analyzes multiple conversations to detect multi-session abuse.
- Anthropic's data retention policy, announced in July, retains user data for 30 days for covered models.
- Covered models include all Mythos-class models and future models with similar capabilities.
- Anthropic generally adheres to Zero Data Retention except for covered models like Fable.
- Anthropic's human review process uses controlled access and tamper-proof logs.
- Anthropic's annualized revenue run rate is $65 billion; investors suggest IPO at $2 trillion; OpenAI is also working on IPO.
Entities
Institutions
- OpenAI
- Anthropic