The artificial-intelligence industry may be entering a new and more secretive phase, one in which the most advanced systems are deemed too dangerous for public release. That shift became dramatically visible in April when AI company Anthropic announced that one of its newest models, Claude Mythos, would not be made broadly available because of the risks it allegedly posed to cybersecurity, public safety, and national security.
According to reporting and analysis published by the scientific journal Nature, Anthropic claimed that Mythos was capable of identifying vulnerabilities across major operating systems and web browsers currently in use. The San Francisco-based company warned that unrestricted deployment of the model could have severe consequences for economies, infrastructure, and governments if malicious actors gained access to its capabilities.
Instead of releasing the system publicly, Anthropic launched a tightly controlled initiative known as Project Glasswing, allowing only around 50 trusted organizations to use the model under supervision. The decision has sparked intense debate inside the global technology and scientific communities over whether AI development is moving away from openness and toward a restricted-access future shaped by corporate control and government oversight.
Experts say the implications could extend far beyond cybersecurity. Helen Toner, interim executive director at Georgetown University’s Center for Security and Emerging Technology and a former OpenAI board member, told Nature that Anthropic’s move is unlikely to remain an isolated case. She suggested it may instead signal the beginning of a broader industry trend in which companies deliberately withhold their most advanced AI systems from the public.
That prediction appeared to gain support almost immediately. Just days after Mythos was announced, OpenAI introduced GPT-5.4-Cyber, a cybersecurity-focused model available only to approved researchers and organizations. A follow-up version, GPT-5.5-Cyber, was also launched under similar restrictions, reinforcing the sense that major AI firms are beginning to treat their most powerful tools less like consumer products and more like sensitive technologies requiring controlled access.
The shift marks a dramatic turning point in the long-running debate between advocates of open and closed AI systems. For years, many researchers argued that transparency and public availability were essential for scientific progress, allowing academics and developers to study, improve, and build upon existing models. Open access, supporters claimed, encouraged innovation while ensuring that AI development remained accountable to the broader public.
Now, however, concerns about misuse are increasingly reshaping that philosophy. Companies developing cutting-edge models fear that unrestricted access could enable cyberattacks, disinformation campaigns, surveillance operations, or even biological threats.
Anthropic’s concerns about Mythos remain difficult for outsiders to independently verify because researchers without access to the system cannot fully evaluate its capabilities. Nevertheless, some cybersecurity experts believe the company’s warnings should be taken seriously. Ciaran Martin, former head of the United Kingdom’s National Cyber Security Centre, described Mythos to Nature as a potentially significant leap forward in AI capability and suggested the pace of development appears to be accelerating rapidly.
Current AI systems already contain so-called guard rails intended to prevent harmful use. These protections usually involve internal instructions designed to stop models from answering dangerous or malicious prompts. But researchers acknowledge that such safeguards are often imperfect and vulnerable to “jailbreaking,” a process in which users manipulate the model into bypassing restrictions.
Supporters of restricted access argue that keeping advanced systems out of public hands buys time for defenders to strengthen security before attackers can exploit the technology. Vasilios Mavroudis, an AI safety researcher at the Alan Turing Institute in London, told Nature that the same AI tools capable of helping cybersecurity professionals identify software weaknesses could also be used by hackers searching for vulnerabilities.
The concerns are not limited to digital security. AI companies are increasingly worried about the possibility that highly specialized systems could aid in the development of biological weapons or other dangerous technologies. In April, OpenAI introduced GPT-Rosalind, a model designed for life-science research, but restricted access to approved users under a “trusted-access” framework. Google has also limited availability of its AI co-scientist platform to selected researchers who apply for entry.
For scientists, the trend raises difficult questions about equality and access. If the most advanced AI systems remain restricted to elite institutions, wealthy corporations, or politically connected researchers, smaller laboratories and developing countries could struggle to keep pace. Researchers already complain that rising subscription costs for public AI systems are deepening inequality within academia and limiting who can participate in cutting-edge research.
The growing power of AI is also drawing increased attention from governments. According to Nature, officials in the administration of US President Donald Trump are reportedly considering forms of formal oversight for advanced AI models, including possible government review procedures before release. At present, oversight in the United States remains largely voluntary, with some companies submitting systems for evaluation to the Center for AI Standards and Innovation at the National Institute of Standards and Technology.
Britain’s AI Security Institute performs a similar role, but it remains unclear how much influence these bodies currently exert over corporate decisions. Anthropic has not publicly stated whether outside evaluations played a role in its choice to restrict Mythos.
Legal experts suggest that governments may eventually treat advanced AI as a “dual-use” technology, meaning software that has both civilian and military applications. Such classifications already apply to weapons systems, surveillance tools, and sensitive technologies subject to export controls and national-security regulations.
If AI models are brought under similar frameworks, governments could potentially regulate cross-border access, restrict exports, and impose rules governing how companies monitor users and store activity logs. Some experts believe parallels already exist with debates surrounding spyware and cyber-intrusion software capable of being used for espionage.
Yet enforcing meaningful controls on AI may prove extraordinarily difficult. Anthropic chief executive Dario Amodei has warned that open-source developers and Chinese AI firms could replicate Mythos-level capabilities within months, potentially undermining attempts by Western governments or companies to maintain restrictions.
At the same time, geopolitical concerns are beginning to shape discussions around AI governance. Nature reported that diplomatic talks between the United States and China in May included discussions on AI guard rails and preventing malicious use by non-state actors, signaling that governments increasingly view advanced AI as both a technological and strategic issue.
The emergence of restricted-access AI marks a profound change in how the world’s most powerful software may be developed and distributed. What was once framed as a democratizing technology is increasingly being treated as a potentially dangerous capability requiring supervision, licensing, and control.
For researchers, governments, and the public, the challenge now lies in determining who gets access to the future’s most powerful AI systems — and who gets left out.
