Agentic AI
,
Artificial Intelligence & Machine Learning
,
Governance & Risk Management
Enterprise Users Face Delayed Access as Trusted Partners Get Early Preview
OpenAI limited its release of GPT-5.6, its newest model with strong capabilities around coding, R&D and cybersecurity, to a short list of users after the Trump administration requested access to the model and the list of users.
See Also: Edge Transformation: Top 5 SASE Predictions and Trends
The staggered rollout for GPT-5.6 came weeks after the government exerted emergency export controls over Anthropic’s Fable 5 model, saying reported jailbreak vulnerabilities could threaten national interests if cyber adversaries got access to Fable’s advance capabilities to uncovering software vulnerabilities. A week before on June 5, Trump issued an executive order asking frontier labs to voluntarily hand over models for safety testing.
The Information reported that OpenAI CEO Sam Altman told employees the Trump administration had requested the company release GPT-5.6 to a short list of partners before a full release.
The company confirmed its decision in a blog post announcing GPT-5.6’s release. Access to the model will start “with a limited preview for a small group of trusted partners whose participating has been shared with the government ahead of [the] launch.” ISMG reached out to OpenAI for a list of these companies but the company didn’t immediately respond to the request. This preview approach is similar to Anthropic’s launch of its Claude Mythos model, voluntarily limiting access to members of Project Glasswing to what it described as a “dangerous” model in the wrong hands.
The executive order provides a voluntary 30-day period for the government to review frontier artificial intelligence models, but the order doesn’t include language related to reviewing or approving customers who can access the models.
While the company said it intends to follow the government’s request, it also wants to make the three tiers of GPT-5.6 – Sol, Terra and Luna – generally available in the next few weeks.
“During this preview, we will continue testing and coordinating closely with partners as we work toward broader availability,” the company said. “We don’t believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders and global partners who need them.”
OpenAI said that it took this step because it’s the “strongest path to broader availability in the coming weeks.”
The company said GPT-5.6 is its strongest model yet, with the flagship Sol Ultra version able to help defenders identify vulnerabilities, understand attack paths and develop fixes. Sol Ultra also provides advanced coding, long-horizon reasoning and sub-agent orchestration, which support large research, coding and engineering projects.
Enterprise AI users agree that regulating AI models is difficult but necessary. But it leaves enterprises in the dark about when they can use a model they’re interested in, forcing them to look to other methods to optimize AI technologies.
Jared Shulman, co-founder and CEO of fintech startup Daylit, told ISMG that on the one hand, the government intervening with models like GPT-5.6 shows how powerful these models really are, even though the current regulatory approach feels reactive.
Many enterprises are choosing to be model agnostic because each model works best for certain tasks, but Shulman said giving a small group of users privileged access essentially gives organizations permission to use that model over others that may be a better fit for their needs.
“Doing a staggered release creates an updated incentives structure and market dynamic where people are jockeying for access to get on that short list,” Shulman said.
Contrast Security co-founder and chief technology officer Jeff Williams said that the government and frontier labs have very little choice in how to balance risk and innovation.
“This buys us a little bit of time to prepare for what could come even from other countries,” Williams said. “But we can’t really nerf the technology, so this is at least a small step.”
Williams said many enterprises have chosen to focus on building harnesses and skills that AI agents can tap into, which often side-steps many limitations of a large language model. This means some enterprises may find they don’t even need to use something as robust as GPT-5.6.
OpenAI said that during the preview period with the first group of users, it will work with enterprise customers on longer-term approaches, including the possibility of adding privacy-preserving detection, customer-operated safety controls and calibrated access.
Shulman said adding identity or permission-based access to models could get complicated, especially in the application or agent layer.
“In practice, models sit on the application layer for enterprises. It gets messy because the question becomes who is using the model in this layer and how are agents tapping these models,” he said.