OpenAI unveils new safety system to prevent misuse of customer data

# OpenAI Unveils Privacy-Focused Safety System as AI Competition Intensifies

OpenAI is introducing a new privacy-focused safety system designed to detect potential misuse of its artificial intelligence models without requiring the company to access or retain customers’ underlying data. The new technology, called Private Safety Processing, is being previewed with a select group of customers as AI companies face growing pressure to balance safety, privacy and security.

The system is designed to identify patterns that could indicate harmful activity across multiple interactions. Instead of simply examining individual requests, an automated agent can look for connections between conversations that may reveal attempts to evade AI safety protections.

For example, OpenAI says the system could identify a user attempting to develop malware for a cyberattack while deliberately spreading related requests across several conversations. Private Safety Processing is designed to detect these patterns without requiring human employees to read the customer’s conversations.

When the system identifies potentially suspicious activity, OpenAI can receive signals indicating that misuse may be taking place. The company could then contact the customer for additional context or work with them to address the issue. Customers would also retain the option of voluntarily sharing relevant data with OpenAI.

OpenAI says the new approach goes beyond its existing Zero Data Retention policy. Under Zero Data Retention, agents used through the OpenAI API can monitor activity for potential abuse on a session-by-session basis without retaining customer data. Private Safety Processing is designed to provide broader analysis while keeping the underlying information protected.

The announcement also comes as OpenAI seeks to differentiate its privacy approach from rival Anthropic. Anthropic announced a policy allowing certain models to retain user sessions and conversations for up to 30 days. The policy applies to its Mythos-class models and future systems with similar capabilities, although Anthropic says human access is restricted to a small group of authorized reviewers operating under controlled procedures.

For businesses handling sensitive information, the difference could become an important factor when choosing an AI provider. Enterprises increasingly want powerful AI systems but are also concerned about confidential information being stored, reviewed or exposed.

The new safety system arrives at an important moment for OpenAI itself. The company is reportedly preparing for a potential initial public offering while continuing to invest heavily in AI infrastructure and product development. OpenAI’s revenue reportedly reached $6.7 billion in the second quarter, up from $5.7 billion in the first quarter, although the company’s operating losses remain significant.

OpenAI is also expanding ChatGPT beyond a traditional chatbot through its “super app” strategy, combining conversational AI with coding capabilities, web automation and tools designed to complete complex tasks.

Meanwhile, Anthropic continues to grow rapidly, reporting $11.6 billion in quarterly revenue and projecting significantly higher annual revenue. The competition between the two companies is therefore extending beyond model performance into areas such as enterprise privacy, safety and trust.

With Private Safety Processing, OpenAI is betting that stronger misuse detection does not have to come at the cost of customer privacy. If successful, the technology could become an important part of how AI companies address the growing challenge of keeping increasingly powerful models safe while protecting the sensitive information of the people and businesses using them.