Histoire
août 20, 2026

OpenAI fait de la confidentialité son arme la plus affûtée contre Anthropic

OpenAI teste une surveillance d’abus inter‑sessions qui, selon l’entreprise, respecte une rétention de données nulle, marquant un contraste appuyé avec les règles de rétention d’Anthropic pour ses modèles les plus puissants. Le duel porte de plus en plus sur la question de savoir si la sûreté exige de voir les données des clients.

The AI safety race is becoming a fight over who gets to see the data. OpenAI is betting that enterprise customers will choose monitoring that can spot sustained misuse without handing their conversations to the model maker.

The divide sharpened in July, when Anthropic introduced a policy allowing it to retain sessions for 30 days on certain “covered models,” including Mythos-class systems and future models of similar capability, for safety analysis. The policy alarmed enterprises handling sensitive information, even as Anthropic said any human review would run through a controlled process involving a small group of approved reviewers and tamper-proof logs.

OpenAI’s answer is now being tested with early customers. Its proposed Private Safety Processing expands the usual per-interaction safeguards of Zero Data Retention into cross-session monitoring: automated systems would look for patterns such as repeated attempts to probe guardrails, coordinated activity across accounts, or a harmful objective scattered over many prompts.

The company’s central promise is that it can detect the pattern without opening the underlying conversations to staff. “OpenAI personnel do not receive access to the customer content even when it is flagged,” it said; instead, the system would return a narrowly defined signal about the type of suspected activity. Customers could then supply context voluntarily if they wanted to contest an enforcement decision or assist an investigation.

That is the crucial contrast. Both labs accept that one-off screening may miss long-running abuse, including an actor distributing malware-related requests across sessions. But Anthropic’s approach permits retention and potentially tightly controlled human review for its most advanced models; OpenAI is presenting encryption, customer-controlled keys and automated signals as a way to preserve privacy while widening the safety net.

OpenAI said it plans to begin rolling out the system and publish a technical white paper in September. Zero retention will not be absolute: apparent child sexual abuse material can still be retained for manual review and legally required reporting.