OpenAI: Surveillance Without Retention in Response to Anthropic's Policy

OpenAI is offering select clients a new automated abuse detection system, described as not retaining any customer data. This approach contrasts with the 30-day data retention policy implemented by Anthropic for certain models, as both labs compete for market share and funding.

Anthropic Valued at $65 Billion, IPO Speculated at $2 Trillion

The numbers are also a battleground. According to a recently published report, Anthropic experienced faster growth than OpenAI in the second quarter. Anthropic's annualized revenue estimate reaches $65 billion. Investors from Anthropic have also stated that it could go public at $2 trillion. OpenAI is also working on its IPO. Both companies are seeking leverage points to differentiate themselves.

OpenAI Introduces Private Safety Processing to Selected Clients

OpenAI has announced a privacy-focused security approach to detect misuse. The company is presenting a new service to selected clients, Private Safety Processing, described as automated and not retaining any customer data. This positioning is highlighted in contrast to Anthropic's recent data retention policy.

ZDR Already Available: Session Monitoring Without Retention

OpenAI already offers Zero Data Retention (ZDR), which relies on agents within its API to monitor abuse at the session level. This framework ensures that customer data is not retained by the company while allowing the detection of suspicious activities without human intervention. Anthropic claims to largely adhere to this practice, with the exception of covered models like Fable.

Anthropic Retains Data for 30 Days for Mythos and Similar Future Models

In July, Anthropic announced a policy allowing for the retention of user data for 30 days for "covered models." This category includes all models in the Mythos class and future models with similar capabilities, according to the company. Anthropic justifies this measure by security imperatives, to filter and analyze potentially inappropriate behaviors. Human review of the data may occur through a controlled access pathway involving a small set of approved reviewers, with each session logged in a tamper-proof manner that reviewers cannot delete or modify. However, this retention has raised significant concerns among companies handling sensitive data that refuse to have it retained or inspected.

Multi-Session Detection Without Human Review, Signals, and Application

OpenAI describes Private Safety Processing as a solution that extends ZDR to allow monitoring over a longer period, taking into account inputs and outputs across multiple exchanges rather than just one. The detection mechanism uses an agent that, when activated, can collect interactions and review them across multiple sessions to identify potential signs of abuse. According to a spokesperson, the tool particularly targets malicious uses that unfold over time, such as when an individual attempts to create malware for a cyberattack by dispersing their requests, all without the user's conversations being reviewed by a human. If the agent is triggered, the system can send a specific signal to OpenAI, allowing the company to determine if intervention is needed. In this case, OpenAI reaches out to the client for additional information or to collaborate on resolving the issue; the client retains the option to share or not share data with OpenAI, based on their own decision. This change comes amid a context where the rise of models increases the risks of abuse, as the demand for protective measures intensifies and providers must balance customer privacy with the need to monitor usage.