OpenAI is taking direct aim at Anthropic’s controversial data-retention policy with a new service that monitors for AI misuse while keeping zero customer data. The company began previewing Private Safety Processing this week for select enterprise customers, pitching it as a privacy-first alternative to the 30-day retention window Anthropic announced in July for its Mythos-class models, including Fable.
The technical approach matters here. Most AI providers, including both OpenAI and Anthropic, already operate under Zero Data Retention, where automated agents scan individual sessions for abuse without storing anything. What makes Private Safety Processing novel is its long-horizon monitoring: the system can connect patterns across multiple conversations, not just one. A malicious actor trying to build malware might spread requests across sessions to avoid triggering per-session alerts. OpenAI’s agent can now catch that distributed behavior without any human ever reading the underlying conversations.
If the system flags something, OpenAI receives only a narrowly defined signal indicating a specific type of activity, not the content itself. The company then decides whether enforcement is warranted and may reach out to the customer for context. Anthropic, by contrast, allows human review of covered model data through a controlled access path with approved reviewers and tamper-proof logs. That’s a meaningful philosophical difference: OpenAI is betting enterprises would rather have an automated system that alerts on patterns than a human review process that touches their data at all.
The timing isn’t coincidental. Anthropic just crossed a reported $65 billion annualized revenue run rate while OpenAI’s Q2 growth reportedly slowed. Both companies are racing toward IPOs, and enterprise trust around data handling has become a genuine competitive battleground. OpenAI clearly sees Anthropic’s retention policy as a vulnerability worth exploiting.
