As artificial intelligence models increase in capability, so too does the potential for their misuse, leading to a strong demand for safety measures to prevent abuse. AI firms are now tasked with balancing their enterprise clients’ privacy concerns while monitoring usage for any potential issues.
In a strategic move to outpace its competitor Anthropic, OpenAI has unveiled a new privacy-focused monitoring solution aimed at preventing misuse. This initiative, called Private Safety Processing, is an automated system that identifies potential abuse without retaining any customer data.
This approach stands in stark contrast to Anthropic’s recently implemented data-retention policy, which has raised concerns among some clients. The policy allows Anthropic to store user data, including past sessions and conversations, for up to 30 days for specific models. These include all Mythos-class models and other future models with similar capabilities.
Introduced in July, Anthropic’s policy was intended to enhance safety by enabling the analysis of potential misuse. However, it has alarmed enterprises that manage sensitive information and prefer not to have their data stored or analyzed by the AI firm.
Like many AI companies, OpenAI provides a level of privacy through its Zero Data Retention policy. This policy utilizes agents within the OpenAI API to monitor for any abusive activity on a per-session basis, ensuring that customer data is not stored while still allowing the company to detect suspicious behavior without human oversight. Notably, Anthropic also primarily follows this Zero Data Retention protocol, with exceptions for certain models.
OpenAI asserts that Private Safety Processing represents an advancement of the Zero Data Retention framework. It functions as a comprehensive safety monitoring system that evaluates the inputs and outputs from multiple conversations rather than just a singular session. The monitoring is autonomously conducted by an agent that, when activated, captures interactions across sessions, looking for indicators of potential misuse.
According to a spokesperson, this technology aids OpenAI in identifying malicious activities that might unfold over several sessions. For instance, a bad actor attempting to create malware for a cyberattack may spread their inquiries over time to evade detection. The Private Safety Processing system is capable of scrutinizing these interactions for abuse signs without the need for human examination of the user’s dialogues.
If the system is triggered, it can send a specific alert to OpenAI regarding particular activities. Based on this alert, OpenAI can determine if any enforcement action is warranted. If so, they may contact the customer for additional context or collaborate with them on the issue, allowing the customer to voluntarily share data if they choose.
In contrast, Anthropic allows for human reviews of customer data, but only via a controlled process involving a limited group of authorized reviewers. Each review session is documented in a secure log that cannot be altered or concealed by the reviewers.
The rivalry between OpenAI and Anthropic is currently intensified, with both companies striving to gain the upper hand. Recent reports indicate that OpenAI’s growth in the second quarter was slower compared to that of Anthropic, which has reportedly achieved an annualized revenue run rate of $65 billion. Investors from Anthropic have suggested that it could achieve a $2 trillion IPO, while OpenAI is also preparing for its own public offering.
