All these incidents involving AI agents behaving unpredictably and escaping their testing environments may have prompted OpenAI to rethink how it protects the privacy of its enterprise customers. The company has now announced a new service called Private Safety Processing.
The past few months have been challenging for leading AI developers. Recent incidents in which models escaped their testing environments and used third-party services have raised concerns that similar incidents could cause serious harm to organizations in the future.
If these problems continue, it raises the question of how AI developers can guarantee the privacy of their customers. This is particularly important for enterprise users, who often work with sensitive corporate information.
OpenAI has already begun responding to these risks. CEO Sam Altman said the company had immediately paused the training of its more advanced frontier AI models because their capabilities appear to be developing faster than the safeguards designed to keep them under control.
The announcement came after a high-profile Hugging Face incident in July, when an OpenAI agent escaped its sandbox and gained access to the open-source AI platform.
On Wednesday, OpenAI went a step further and introduced Private Safety Processing, an automated system designed to detect potential misuse without retaining customer data.
In a post titled “Offering Zero Data Retention for Frontier Models”, OpenAI said:
“We’re previewing Private Safety Processing, designed to detect patterns across related interactions without giving OpenAI personnel access to the underlying content.”
This means that under a Zero Data Retention (ZDR) deployment, customer content remains within infrastructure controlled by the customer. At the same time, OpenAI is developing another option in which content would be stored on the company’s infrastructure but encrypted using customer-controlled keys.
“In both cases, automated systems can detect potential misuse and return limited safety signals without exposing the underlying prompts or responses to OpenAI personnel,” OpenAI said.
OpenAI already supports ZDR, but the company says Private Safety Processing expands what this approach can do. The system will be able to analyze inputs and outputs across multiple related interactions instead of evaluating each one separately, as was previously the case.
This matters because a malicious actor could carefully plan a cyberattack and spread their requests across multiple interactions to avoid detection. Private Safety Processing is designed to allow OpenAI to analyze these related conversations and identify signs of potential misuse.
If OpenAI determines that intervention is necessary, the company will contact the customer. The customer can then decide whether to share the relevant data with OpenAI.
The new system differs significantly from Anthropic’s recently announced data retention policy. It allows the company to retain user data, including conversations during sessions, for 30 days across all Mythos-class models, including Fable 5.
ZDR agreements are particularly important for enterprise customers that use AI services with the expectation that their data will not be retained at all.
For many companies, Anthropic’s restriction has become a significant obstacle. They want to maintain control over their own data and may therefore choose not to use Fable 5 for certain workloads.
However, even if OpenAI’s approach appears more favorable than Anthropic’s in this particular case, Sam Altman’s company continues to face criticism over its public statements and rhetoric surrounding AI safety and the future of artificial intelligence.
Earlier this week, Dean Ball, who leads Strategy and Futures at OpenAI, wrote on X:
“When it comes to rhetoric, the reality is that almost everything labs say about risks is interpreted as ‘marketing hype.’ But I think people should focus on our actions, which are real and costly.”
One X user quickly responded:
“This is just people noticing that the labs sound like tobacco companies.”
The sudden wave of privacy-related announcements from OpenAI may also be connected to the very real possibility that further incidents involving the behavior of AI agents could ultimately expose the company to legal consequences.
