Zero retention vs data logs: OpenAI and Anthropic take different safety routes to detect harmful activity

OpenAI announced to test a new zero-retention security technique, allowing businesses to use its most advanced AI models without having to share any customers’ data.  The technique is being called “Private Safety Processing,” which is currently being tested among early customers, and it could be useful for businesses handling sensitive information.

On the other hand, Anthropic is taking a completely opposite road with a 30-day data-retention policy. During this period, businesses using its Fable 5 and Mythos 5 models will have to share data, which Anthropic says is necessary for security.

Must read: Google tops AI trust rankings, OpenAI and Anthropic struggles to keep up

OpenAI’s Zero Data Retention

According to the company blog post, the zero-retention security technique promises businesses that OpenAI will not retain their prompts or model responses, and customer content will not be available to OpenAI personnel for review.

The technique is designed to identify suspicious patterns across multiple related interactions without actually storing the users’ data. Therefore, the system maintains security and detects misuse without sacrificing the privacy protections of zero data retention. ZDR uses agents within the OpenAI API that track activity while keeping customer data protected, but companies can still scan for bad activity without human intervention.

When the system detects any unusual activity, it generates a limited signal describing what kind of risk was detected. OpenAI can use that limited information to decide whether it needs to take action. Businesses can investigate the alerts and also choose to send relevant information to OpenAI.

Must read: OpenAI to tighten AI safety rules after AI agents hacked startup; What’s changing

Anthropic’s 30-day retention: Why its going in a different direction?

Recently, Anthropic announced 30-day data retention on its most capable models, which claims to collect customer data to identify threats. The company considers it essential to detect and prevent sophisticated attacks that span multiple requests.

Anthropic in a risk report said, “We have recently announced our plan to require 30-day data retention on our most capable models—a decision we believe 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 which we believe is essential to detect and prevent sophisticated attacks that span multiple requests.”

OpenAI vs Anthropic

Both OpenAI and Anthropic are taking different approaches to AI safety. OpenAI has reportedly paused some model-training work to focus on safety, and it has also had a two-week pause in reinforcement-learning training for frontier models. Whereas Anthropic says it currently does not believe a similar pause is necessary.