OpenAI Plans Limited Rollout of New AI Model Amid Cybersecurity Concerns

OpenAI is preparing to release a new artificial intelligence model with advanced cybersecurity capabilities through a restricted rollout, limiting early access to a select group of companies rather than making the system publicly available, according to a source familiar with the matter.

The move reflects growing concern across the AI industry that increasingly capable models may be powerful enough to enable serious cyberattacks if broadly distributed without safeguards.

OpenAI Expected to Mirror Anthropic’s Restricted Release Strategy

The planned launch would follow a similar strategy adopted by Anthropic, which recently announced that its new Mythos Preview model would be available only to a curated group of technology and cybersecurity firms because of concerns over the model’s advanced hacking capabilities.

If OpenAI proceeds with the limited rollout, it would mark another major sign that top AI developers are becoming more cautious about releasing frontier models with offensive cyber potential.

Why AI Companies Are Restricting Advanced Cyber Models

Executives and security officials have increasingly warned that the latest generation of AI systems may be reaching a threshold where they can autonomously identify vulnerabilities, generate exploits, and assist in sophisticated cyber operations.

Former U.S. government officials and cybersecurity leaders have spent the past year cautioning that AI could eventually be used to disrupt critical infrastructure, including:

Potential High-Risk Targets

• Water treatment systems
• Electrical grids
• Financial institutions
• Other critical public infrastructure

Those concerns appear to be moving from theoretical to immediate, according to security experts monitoring the rapid evolution of AI-assisted hacking tools.

OpenAI’s Trusted Access for Cyber Program Laid the Groundwork

OpenAI first signaled its cautious approach in February when it introduced its Trusted Access for Cyber pilot program following the release of GPT-5.3-Codex, the company’s most cyber-capable reasoning model to date.

The invite-only initiative provides approved organizations with access to more advanced cybersecurity-focused AI systems intended to support legitimate defensive research and threat analysis.

At launch, OpenAI committed $10 million in API credits to participating organizations.

Security Experts Say the Industry Has Reached a Turning Point

Cybersecurity leaders say restricting access may slow the spread of dangerous capabilities, but it will not stop them entirely.

“You can’t stop models from doing code enumeration or finding flaws in older codebases,” said Rob T. Lee. “That capability exists now.”

According to experts, similar systems with comparable offensive capabilities are likely to emerge publicly within weeks or months, regardless of individual companies’ release strategies.

Wendi Whitmore said during a panel at the HumanX conference in San Francisco that the pace of development means highly capable cyber models are unlikely to remain restricted for long.

Meanwhile, Adam Meyers described Anthropic’s Mythos capabilities as “a wake-up call” for the cybersecurity industry.

Limited AI Releases Resemble Responsible Vulnerability Disclosure

Some experts argue that staggered AI releases mirror how the cybersecurity industry has long handled software vulnerabilities.

Under traditional responsible disclosure practices, researchers privately share serious security flaws with affected vendors before broader public release, allowing time for mitigations and patches.

“That’s the same debate we’ve had for decades around responsible vulnerability disclosure,” Lee said.

Security analysts say this approach makes the most sense when companies are concerned about a model’s ability to generate new exploits, rather than simply identify vulnerabilities.

Unclear Whether OpenAI Will Expand Access Later

It remains uncertain whether OpenAI plans to eventually release the forthcoming model more broadly.

Anthropic has stated it does not currently intend to make Mythos Preview publicly available, though the company has indicated future Mythos variants could launch more widely if sufficient safety guardrails are developed.

Existing AI Models Already Show Similar Capabilities

Even with tighter restrictions, some researchers argue the gap between highly restricted frontier models and publicly accessible AI tools may already be narrowing.

Researchers at security firm Aisle said this week that widely available AI models can already identify certain vulnerabilities and exploits similar to those discovered by Anthropic’s Mythos.

That suggests the cybersecurity risks associated with advanced AI may no longer be confined to experimental or tightly controlled systems.

AI Security Debate Intensifies as Model Capabilities Advance

OpenAI’s reported decision to limit release of its next cybersecurity-focused model underscores a broader shift in the AI industry: developers are becoming increasingly wary of the unintended consequences of their most powerful systems.

While restricted access may delay widespread misuse, experts say the broader challenge remains unresolved as AI capabilities continue to accelerate.

The debate is no longer whether advanced AI can assist in cyber operations—it is how quickly those capabilities will become broadly accessible.

William Faulkner

“Social media evangelist. Baconaholic. Devoted reader. Twitter scholar. Avid coffee trailblazer.”