Anthropic has suspended access to its most advanced AI models for a large portion of its users following intervention by the U.S. government. However, a limited group of organizations appears to retain access to Mythos Preview for the time being, an experimental version of the model used for cybersecurity research.
Last week, the U.S. government ordered Anthropic to block the distribution of certain advanced AI models to foreign users without a license. As a result, the company withdrew Fable 5 from public availability and halted the further rollout of Mythos 5.
According to Bloomberg, however, this restriction does not apply to all users. A number of organizations participating in Project Glasswing, a closed program by Anthropic focused on cybersecurity research, can still use Mythos Preview.
Project Glasswing consists of approximately 200 organizations selected by Anthropic to investigate the capabilities and risks of advanced AI systems. The model is used to analyze software, networks, and systems for vulnerabilities.
Participants include AWS, Cisco, and JPMorgan Chase, among others. Industrial cybersecurity specialist Dragos also confirmed that it still has access to Mythos Preview.
The preview version attracted attention earlier this year because the model identified security vulnerabilities during large-scale tests. According to those involved, the system discovered vulnerabilities in a wide range of software platforms, web browsers, and operating systems. It was precisely because of these capabilities that Anthropic decided to severely restrict access.
Uncertainty Regarding the Scope of the Ban
Following the U.S. decision, there was initially confusion within the Glasswing community about exactly which models were subject to the restrictions. The government measure did not explicitly mention Mythos Preview, while Anthropic primarily referred to Fable 5 and Mythos 5 in its public communications.
As a result, it remained unclear whether research versions of the model would also be removed. It now appears that at least some participants have retained their access.
Anthropic has not provided any further clarification on this matter.
European participants affected
Not all organizations within the program benefit from this exception. The European cyber agency ENISA had previously received an invitation to participate in Project Glasswing. However, it was notified last week that access is no longer possible.
This suggests that Anthropic assesses access on a per-organization basis, taking U.S. export restrictions into account. The exact criteria used are unknown.
The situation underscores how quickly advanced AI systems are taking on a geopolitical dimension. While AI models were initially viewed primarily as tools for productivity and software development, governments are now increasingly regarding models capable of autonomously detecting vulnerabilities as technology with strategic and security-related applications. This also increases the likelihood that access to such systems will be further regulated in the future.