1 week ago
OpenAI Tests Zero-Retention Safety as Anthropic Retains Data
OpenAI is testing a new way to find unsafe activity when businesses use its powerful AI models.
The system looks at patterns across several related interactions instead of examining only one question and answer.
OpenAI says it can do this without keeping the business’s prompts or the AI’s responses.
It would receive only a small signal explaining what kind of risk was found.
Business customers can investigate alerts and may choose to share more information.
Customer data can stay on the customer’s own systems or be encrypted with keys controlled by the customer.
Anthropic is choosing a different approach for its most capable models.
It plans to keep business data for 30 days because it says this helps detect attacks spread across multiple requests.
These OpenAI controls are for eligible enterprise and API customers, not Free, Plus, Go, or Pro ChatGPT users.
OpenAI is testing Private Safety Processing with selected enterprise and API customers.
The system aims to detect harmful activity across related interactions without retaining prompts or model responses.
OpenAI says it would receive only a limited signal describing the detected risk, while customer data remains protected.
Anthropic says 30-day retention on its most capable models is necessary to detect sophisticated attacks spanning multiple requests.
OpenAI plans a broader rollout and technical white paper in September, while consumer ChatGPT settings remain unchanged.
- Who
- OpenAI and Anthropic, along with their enterprise, business, and API customers.
- What
- OpenAI is testing a zero-data-retention safety system, while Anthropic plans 30-day retention for its most capable models.
- Where
- The measures apply to business and API services, with OpenAI customer data kept on customer-controlled infrastructure or protected using customer-controlled encryption keys.
- When
- OpenAI is testing the system now and plans a technical white paper in September; Anthropic announced its retention plan recently.
- Why
- OpenAI aims to detect misuse across multiple interactions without retaining customer content, while Anthropic says retaining data is needed to detect sophisticated attacks.
OpenAI’s Zero-Retention Approach
Anthropic’s 30-Day Retention Approach
Detecting multi-step misuse
OpenAI’s Zero-Retention Approach
OpenAI is testing agents that identify suspicious patterns across related interactions without storing the underlying prompts or responses.
Anthropic’s 30-Day Retention Approach
Anthropic says retaining data for 30 days is essential to detect and prevent sophisticated attacks spanning multiple requests.
Privacy versus security
OpenAI’s Zero-Retention Approach
OpenAI’s approach is intended to preserve zero data retention and prevent customer content from being available to OpenAI personnel for review.
Anthropic’s 30-Day Retention Approach
Anthropic acknowledges that retention may be unpopular and could create business risks, but considers it necessary for security.
Scope of the measures
OpenAI’s Zero-Retention Approach
OpenAI’s system is being developed for eligible enterprise and API customers, not consumer ChatGPT subscription plans.
Anthropic’s 30-Day Retention Approach
Anthropic’s announced retention requirement applies to business customers using its most capable models.
Key facts
- OpenAI system
- Private Safety Processing is being tested with early enterprise and API customers.
- OpenAI data policy
- The system is designed to preserve zero data retention for customer prompts and model responses.
- Safety signal
- OpenAI says it would receive a narrowly defined signal describing the detected risk.
- Customer control
- Data may remain on customer infrastructure or be stored with encryption keys controlled by the customer.
- Anthropic policy
- Anthropic plans to require 30-day data retention on its most capable models.
- OpenAI rollout
- OpenAI plans a broader rollout and a technical white paper in September.
- Consumer coverage
- OpenAI’s new controls do not apply to Free, Plus, Go, or Pro ChatGPT users; their existing settings remain unchanged.
Quotes
Anthropic risk report author
Representative of Anthropic’s risk management team
“We have recently announced our plan to require 30-day data retention on our most capable models—a decision we believe will be unpopular with customers who have come to expect zero retention, and pose real risks to our business success (especially if competitors do not follow), but which we believe is essential to detect and prevent sophisticated attacks that span multiple requests.”
businesstoday.in
firstpost.com
“"We do not retain prompts or responses, and customer content will not be available to OpenAI personnel for review."”
businesstoday.in
Aleah Houze
Head of Product Policy at OpenAI
“We're seeing with more capable frontier models that often risks are emerging not just by looking at one single prompt and response pair, but when you look over time at multiple interactions.”
firstpost.com
Sources
Related news
OpenAI Previews Privacy-Preserving System to Monitor AI Misuse
Anthropic Changes Enterprise Data Retention Rules With New Safeguards
Anthropic Releases Fable 5.1 and Mythos 5.1
OpenAI Tests Private Safety Processing For Paid Customers
AI Labs Score Poorly on Rogue-Model Control Measures
OpenAI Tightens AI Model Security After Hugging Face Incident
Anthropic launches Claude Fable and Mythos 5.1 models
OpenAI’s Astra Reaches Critical Cybersecurity Threshold Ahead of Limited Rollout
Google Prepares Gemini 3.8 Flash to Challenge AI Rivals
