As AI systems have grown in capability, questions about their potential for exploitation have intensified alongside calls for protective oversight mechanisms to prevent such outcomes. AI providers face the ongoing challenge of safeguarding enterprise client data while simultaneously monitoring platform activity for potential irregularities.
The company outlined a user-privacy-focused framework for detecting potential platform abuse. The firm is offering a limited preview of a feature it has named Private Safety Processing to a group of chosen customers. The platform operates by flagging potentially harmful activity in real time, without storing any personal information belonging to users.
The approach conflicts with Anthropic's data retention guidelines, which the company outlined in a recent announcement. The rule, which has drawn complaints from a portion of its user base, permits the AI laboratory to retain account data — including session histories and associated conversations — for up to 30 days across applicable models. According to the firm, the affected lineup covers all Mythos-class products along with forthcoming releases of comparable capability.
Introduced in July, the measure aimed to enable the organisation to review and assess potential misconduct through a structured safety framework. Nevertheless, certain organisations managing substantial volumes of confidential information have raised objections, citing concerns that the AI company could store or access that data.
OpenAI, along with most other companies in the artificial intelligence sector, maintains customer data protections through a framework called Zero Data Retention. To detect misuse on a per-session basis, ZDR deploys agents through the OpenAI API. User data is not stored by the organisation, yet automated systems can still detect irregular activity without requiring manual oversight.
Anthropic generally follows Zero Data Retention protocols, with the exception of its covered models. According to the company, Private Safety Processing represents a newly developed capability that broadens the reach of ZDR. The system characterises this capability as extended-horizon safety oversight, evaluating data across multiple conversation threads rather than treating each exchange in isolation.
An automated agent oversees the process, intercepting exchanges when activated and reviewing them across multiple sessions to detect potential patterns of misuse. According to a company spokesperson, the updated system enables OpenAI to identify harmful AI activity spanning across multiple interactions.
Someone seeking to deploy malicious code for a network intrusion might distribute their queries across multiple sessions to evade detection systems. The system is designed to assess multiple exchanges for indicators of harmful behaviour without requiring direct access to individual user communications.
According to the company, activating the system would result in a tightly scoped notification being transmitted to OpenAI about a particular category of behaviour. Using that indicator, OpenAI holds the authority to determine whether action needs to be taken. According to the spokesperson, OpenAI would contact the client for additional context or to address the matter collaboratively, with the client retaining the option to share data at their own discretion.
According to Anthropic's privacy documentation, any examination of user data by personnel is permitted solely via a restricted pathway limited to a designated group of vetted individuals. According to the firm, each of those review sessions is captured in an audit trail that cannot be altered or deleted by those conducting the reviews.
OpenAI and Anthropic are actively competing for position in the AI industry, with each organisation working to strengthen its standing relative to the other. According to a newly published report, OpenAI's second-quarter revenue expansion trailed the pace recorded by Anthropic. According to reports, Anthropic's yearly revenue run rate has reached approximately $65 billion. Anthropic backers have floated a potential IPO valuation of $2 trillion for the company, while OpenAI is separately pursuing its own public listing.
