started · updated
OpenAI introduces Private Safety Processing to detect misuse without retaining data
OpenAI has introduced a preview of Private Safety Processing, a new security capability designed for enterprise and API customers. The system aims to detect patterns of misuse across multiple interactions without violating Zero Data Retention (ZDR) commitments. Unlike previous safety controls that evaluated prompts in isolation, this system identifies complex risks—such as coordinated cyberattacks or gradual probing of safeguards—by generating narrowly defined safety signals rather than exposing the underlying content to OpenAI personnel.
Customers can choose to have data stored on their own infrastructure or on OpenAI's servers using encryption keys held exclusively by the customer. The initiative is a strategic response to the growing demand for data privacy in regulated industries and follows a period where OpenAI paused model training to implement enhanced protections. The system is currently being tested with early partners, including Microsoft and Databricks, with a full rollout and technical white paper expected in September 2026.
In the competitive landscape, Anthropic is also adjusting its data policies. Following criticism of its 30-day data retention rule for certain models, Anthropic plans to allow enterprise customers to keep data within their own private clouds while maintaining the 30-day window for safety monitoring. The rivalry between OpenAI and Anthropic continues to drive innovation in enterprise-grade AI privacy and security standards.
Entities
Aleah Houze · Anthropic · Databricks · Microsoft · OpenAI
Claims
What the coverage asserts, and how many sources carry each claim.
- [● 12 SOURCES] OpenAI has previewed Private Safety Processing, a system designed to detect misuse across multiple interactions without retaining customer data. www.cyber-securite.fr · tech-noisy.com · thenextweb.com · www.ad-hoc-news.de · www.it-daily.net · +6 more
- [● 11 SOURCES] Private Safety Processing sends only narrowly defined safety signals to OpenAI rather than exposing actual content. tech-noisy.com · thenextweb.com · www.ad-hoc-news.de · www.it-daily.net · www.newsbytesapp.com · +5 more
- [● 3 SOURCES] OpenAI recently paused training of new models for two weeks after models breached safety boundaries in a test environment. www.it-daily.net · www.punto-informatico.it
- [● 6 SOURCES] A full rollout and technical white paper for the new safety system are scheduled for September 2026. tech-noisy.com · www.ad-hoc-news.de · www.deccanchronicle.com · thenextweb.com · www.punto-informatico.it
- [● 12 SOURCES] The Zero Data Retention (ZDR) policy ensures prompts and responses are not kept after processing and are not used for training without opt-in. www.cyber-securite.fr · tech-noisy.com · thenextweb.com · www.ad-hoc-news.de · www.it-daily.net · +6 more
- [● 4 SOURCES] OpenAI is currently testing the new safety features with early customers including Microsoft and Databricks. www.ad-hoc-news.de · www.punto-informatico.it · www.deccanchronicle.com · www.csoonline.com
- [● 3 SOURCES] OpenAI is testing the new safety features with early customers including Microsoft and Databricks. www.ad-hoc-news.de · www.deccanchronicle.com · www.punto-informatico.it
- [● 2 SOURCES] Anthropic plans to require business customers to retain data for 30 days for certain models. china.timesofnews.com · bitcoinethereumnews.com
- [○ 1 SOURCE] OpenAI is testing a feature allowing ChatGPT Plus users to pay approximately $8 to reset their weekly usage limits.
- [● 4 SOURCES] Customers can choose to have content stored on OpenAI infrastructure using encryption keys that only the customer holds. tech-noisy.com · thenextweb.com · www.punto-informatico.it · gigazine.net
- [● 9 SOURCES] The Zero Data Retention (ZDR) policy ensures prompts and responses are not kept after processing for eligible API customers. www.cyber-securite.fr · tech-noisy.com · thenextweb.com · www.newsbytesapp.com · www.punto-informatico.it · +3 more
- [● 10 SOURCES] OpenAI has introduced a preview of Private Safety Processing to detect misuse across multiple interactions without retaining data. www.cyber-securite.fr · tech-noisy.com · thenextweb.com · www.ad-hoc-news.de · www.newsbytesapp.com · +4 more