OpenAI Announces Zero Data Retention, Forcing Anthropic's Privacy Countermove in 48 Hours
新智元 · wechat · 2026-08-30
OpenAI and Anthropic's autumn privacy war: Since June 9, Anthropic required enterprise customers using its strongest tier models (Fable5, Mythos5, etc.) to accept 30-day retention of prompts and outputs, arguing attacks like Best-of-N jailbreaks and data exfiltration are only visible across many requests. On August 19, OpenAI previewed Private SafetyProcessing, promising zero retention — content is discarded after processing, and OpenAI only receives a narrow alert with category and severity, never raw content.
Anthropic's forced response: Per sources, the 30-day retention obligation stays, but data can now sit in the customer's own cloud, with keys and audit under customer control; Claude Code creator Boris Cherny said a truly no-retention option arrives this fall. OpenAI's approach moves the algorithm to the data; Anthropic's shifts custody to the customer.
Consumers are a different story: Both companies' "no retention" applies only to enterprises. Anthropic's consumer chats are retained by default; if training is allowed, de-identified records can stay in training pipelines up to 5 years, policy-violating chats 2 years, risk scores 7 years — with the training toggle defaulting to on since August 2025. OpenAI's personal training toggle also defaults on. The one shared belief: security monitoring is non-negotiable; the only question is whether it requires users' raw data.
More from Safety
- Google Paper: Autonomous AI Research Hallucinates 90% Without Checks — rohanpaul_ai · 2026-09-01
- On token layers and consciousness in RLHF — voooooogel · 2026-09-01
- Agents can't verify people: data enrichment APIs are failing — Dry_Steak30 · 2026-09-01
- Deploying models requires tapping into different reward expectations — FioraStarlight · 2026-09-01
- Open Source Resource for Model Distillation Attacks Shared — k7agar · 2026-09-01
- Technical Critique of OpenAI Safety Report: SSRF Flaw and Anthropomorphism — AlexTensor · 2026-09-01