OpenAI Challenges Anthropic With Zero Retention AI Safety Monitoring
OpenAI previews Private Safety Processing, tracking cross-session AI abuse without retaining customer data, as Anthropic keeps covered-model sessions 30 days.
Summary
On August 19, 2026, OpenAI previewed Private Safety Processing for select customers, an automated system that monitors inputs and outputs across multiple conversations while retaining none of the customer’s data or requiring human review. It expands Zero Data Retention, whose agents monitor individual OpenAI API sessions, to detect prolonged misuse such as malware development split across conversations to evade safeguards.
Anthropic’s July policy retains every session and conversation involving covered models for 30 days so it can investigate potential abuse. Covered models include all Mythos-class systems, Fable, and future models with similar capabilities. Anthropic otherwise largely follows Zero Data Retention, but its exception has alarmed enterprises handling sensitive information. Human inspection can occur through a controlled path restricted to approved reviewers, with each review recorded in a tamper-proof log that reviewers cannot alter or suppress.
When triggered, OpenAI’s system can send the company a narrowly defined signal identifying suspected activity. OpenAI then decides whether enforcement is needed and may contact the customer for context; customers choose whether to share underlying data. The privacy contest comes as OpenAI’s Q2 growth reportedly trailed Anthropic, whose annualized revenue run rate has reached $65 billion. Anthropic investors have discussed a potential $2 trillion IPO, while OpenAI is also preparing an IPO.
Positives
- Private Safety Processing monitors abuse across conversations while retaining none of the customer’s data and avoiding human review.
- Cross-session analysis can detect malware development or other malicious activity deliberately divided among conversations to evade per-session safeguards.
- OpenAI receives only a narrowly defined signal when suspected activity triggers the system.
- Customers may provide underlying data for an enforcement inquiry at their own discretion.
- Anthropic limits human access to approved reviewers and records every review in a tamper-proof log.
Risks & concerns
- Anthropic retains all sessions and conversations involving covered models for 30 days, including Mythos-class systems and Fable.
- The 30-day policy has concerned enterprises that process sensitive information and oppose storage or inspection by Anthropic.
- Private Safety Processing is initially available only as a preview for select OpenAI customers.
- OpenAI can consider enforcement after receiving a warning signal, requiring subsequent contact with the affected customer.
- OpenAI’s Q2 growth reportedly lagged Anthropic as both companies pursue public offerings.