Bottom line: OpenAI is temporarily throttling scaling and RL training and has announced Zero Data Retention for select API customers starting in September, with analysts framing this primarily as IPO preparation and noting that the exact scope of the commitments still needs clarification.
OpenAI announced several measures this week aimed at addressing safety and privacy concerns: a temporary slowdown in scaling pace, a two-week pause in reinforcement learning, and Zero Data Retention for select API customers starting in September. Analysts see this primarily as preparation for a possible IPO.
In an initial statement on Tuesday, OpenAI explained that the company had “temporarily” reduced the pace of its scaling and paused reinforcement learning training. During this time, the company said it had hardened its own research environment, conducted red-teaming, and expanded monitoring. According to OpenAI, the largest planned frontier RL training run remains suspended, while smaller training and evaluation runs are being carried out to assess model behavior, validate safety mechanisms, and gather more evidence for alignment before the large run resumes. The company also announced that it will require stronger evidence of aligned behavior throughout the entire training process going forward. As further measures, OpenAI cited workload isolation, network isolation, and continuous security testing. According to the company, the newly proposed monitoring system is expected to cause a compute overhead of around 20 percent of monitored inference compute, though costs will vary significantly depending on the training and evaluation workload. OpenAI plans to provide further details in a separate blog post.
On Wednesday came the announcement of Zero Data Retention for “eligible API customers.” OpenAI did not specify what criteria determine eligibility – this is only to be explained starting in September in a technical whitepaper. What matters for security leaders at enterprises is that the scope of this commitment depends heavily on the specific contractual relationship. A large share of OpenAI’s revenue does not come directly from enterprise customers but flows through partners such as Microsoft or AWS. If a company uses a tool like Amazon Kiro, for example, which integrates the OpenAI API in the background, Amazon is formally the customer – not the company itself. To obtain Zero Data Retention protection in such a case, a company would need to provide its own API key and would thereby itself become a direct contracting party with OpenAI, gaining corresponding access to so-called lockbox functions. In this scenario, AWS would lose access to the data as well as the associated revenue share.
Consultant Brian Levine, Executive Director at FormerGov, also pointed to the practical complexity behind the promise. OpenAI states that it can monitor for abuse across interactions without employees viewing the underlying content. This is a strong technical commitment, and its actual implementation – for instance, in automated anomaly detection without human review – will need to be examined in detail once the announced whitepaper is available.
Analysts are critical of the timing of the announcements. Carmi Levy, an independent technology analyst, described the steps as a PR move in light of growing safety concerns around agentic AI systems; a two-week pause without regulatory pressure, he said, is “little more than a facade.” Jason Andersen, Principal Analyst at Moor Insights & Strategy, spoke of “pragmatic theater” ahead of an IPO, but also sees a strategic necessity: the sustained success of major AI providers after going public depends on gaining deeper access to enterprise structures – and that requires addressing enterprise customers’ security and risk concerns. For CISOs, this means in practical terms that commitments regarding Zero Data Retention and changed training behavior should only be contractually anchored and verified once the announced technical whitepapers and blog posts are published in September, particularly with regard to indirect usage paths via cloud partners.
Source: www.csoonline.com · Published August 20, 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.8.3.