Anthropic and OpenAI are reshaping who gets to wield the most powerful artificial intelligence models by controlling access rather than releasing them freely. The strategy amounts to a new form of competitive advantage in cybersecurity, where the AI companies themselves decide which researchers, vendors, and government agencies can tap into capabilities that could detect breaches, hunt malware, or reverse engineer attacks.
On Tuesday, Anthropic unveiled Fable 5, a restricted version of its advanced Mythos class of models available to the general public. The model comes with built-in safeguards that block requests about high-risk cybersecurity and biology work, routing those users instead to Claude Opus 4.8, a different model. The company acknowledged it is being deliberately conservative with the initial rollout, meaning some legitimate security researchers may find their requests blocked.
Simultaneously, Anthropic is upgrading users in its restricted Mythos Preview program to access Mythos 5, a less constrained version. The company is also building a formal trusted-access program that will determine which organizations get the unrestricted models, though no launch date has been announced. Over the past two months, organizations have lobbied hard for preview access. Last week, Anthropic expanded the program to more than 150 companies and governments.
OpenAI is already operating a similar two-tier system. The company has been vetting security researchers and organizations to grant access to versions of GPT-5.5 with fewer guardrails, enabling those defenders to hunt for bugs and study attacks without the restrictions placed on mainstream users.
The shift marks a fundamental change in how cybersecurity advantage flows. For decades, the competitive edge came from talent, data, and infrastructure. Now access to cutting-edge AI models is joining that list. Security vendors and researchers desperate to integrate frontier AI into their products and workflows are scrambling to gain entry into these restricted programs.
The approach gives both companies what they need. They can ensure powerful hacking capabilities remain in the hands of defenders while monetizing models that are growing increasingly sophisticated. Both labs are also considering going public, making the revenue potential from tiered access particularly attractive.
The real test will come when trusted-access users begin discovering vulnerabilities and building products that organizations without access cannot match. If the gap widens, it could reshape the entire cybersecurity industry, creating a new divide between those connected to AI giants and those forced to rely on older tools.
Author James Rodriguez: "Anthropic and OpenAI are essentially privatizing the future of cyber defense, and that level of power concentrated in two companies should unsettle anyone worried about fair competition."
Comments