A privacy fight inside enterprise AI safety

OpenAI is previewing a service called Private Safety Processing for selected customers, positioning it as a way to detect abuse of its models without retaining customer data. The move directly contrasts with Anthropic’s recently announced retention policy for some advanced models and shows how enterprise AI competition is shifting from model performance alone to trust, governance, and data handling.

As AI systems become more capable, vendors face a difficult trade-off: they must identify misuse, such as attempts to support cyberattacks, while convincing companies that sensitive business data will not be stored or inspected unnecessarily. For enterprise buyers, privacy is becoming part of the safety feature set.

What OpenAI says the new system changes

OpenAI already offers customers a privacy approach known as Zero Data Retention, or ZDR. In plain terms, ZDR allows automated agents inside the OpenAI API to check for abuse within a session, while the company does not retain the customer’s data.

Private Safety Processing expands that model from single-session checks to what OpenAI describes as long-horizon safety monitoring. Instead of looking only at one conversation, the system can assess inputs and outputs across multiple conversations. If triggered, it analyzes cross-session activity for signs of misuse, still without human review of the user’s conversations.

OpenAI told TechCrunch this is meant to catch malicious behavior that may be deliberately spread over multiple sessions. A hypothetical bad actor attempting to engineer malware for a cyberattack could divide requests to avoid detection. The new system is designed to identify that pattern while sending OpenAI only a “narrowly defined signal” about a specific type of activity.

The reported process is limited:

• automated monitoring across sessions;
• a narrow signal sent to OpenAI if triggered;
• OpenAI decides whether enforcement is needed;
• if more context is required, OpenAI contacts the customer;
• the customer may choose whether to share data.

The key claim is that OpenAI receives a signal, not the full customer conversation.

Why Anthropic is the comparison point

Anthropic announced in July that it may retain user data for 30 days for “covered models,” including all Mythos-class models and future models with similar capabilities. The company says the policy is for safety, allowing it to examine possible misuse. TechCrunch notes that Fable is among the covered models.

That policy has concerned some enterprise customers, especially those handling large volumes of sensitive information. Their objection is not only that data may be stored, but also that it could be inspected by the AI lab.

Anthropic says human review can happen only through a controlled access path involving a small number of approved reviewers. It also says every such review session is recorded in a tamper-proof log that reviewers cannot suppress or modify. Anthropic is therefore emphasizing controlled access and auditability, while OpenAI is emphasizing non-retention and automation.

The business stakes

The timing matters. Competition between OpenAI and Anthropic is intense. TechCrunch cites a report showing OpenAI’s second-quarter growth was slower than Anthropic’s, while Anthropic’s annualized revenue run rate is reportedly $65 billion. Anthropic investors have said the company could go public at a $2 trillion valuation, and OpenAI is also working toward an IPO.

Those figures underline why privacy controls are now strategic. Enterprise customers are not only buying model capability; they are buying assurances about how prompts, outputs, logs, and review processes are handled. In sectors such as finance, healthcare, law, and cybersecurity, contractual privacy language can be as important as benchmark results.

What comes next

Private Safety Processing suggests that AI safety monitoring is becoming a product feature rather than a hidden backend function. Customers will likely ask more detailed questions: which models qualify for zero retention, what signals can be reported, when humans can review data, and what audit trails exist.

The broader direction is clear: powerful models will require stronger misuse detection, but enterprise buyers will pressure vendors to perform that monitoring with less retained data and fewer human touchpoints. OpenAI and Anthropic are taking different routes, and the market will test which balance customers trust most.