Some unauthorized users have gained access to Mythos, the latest artificial intelligence (AI) model from Anthropic that was released on a limited basis due to concerns over high-risk cyberattack potential.
According to Bloomberg on April 21, a small number of users accessed and used the Mythos model through private online communities. The access reportedly occurred on April 7, the same day the company announced the model’s limited release.
The users succeeded in accessing the model by estimating its location using access permissions from third-party partners and publicly available online information.
They continued to use the model afterward, but Bloomberg added that it appears to have been used for general functionality testing rather than for cyberattack purposes.
Mythos is considered capable of not only identifying vulnerabilities in major operating systems and web browsers but also exploiting them to carry out actual hacking, leading the company to allow testing only for a limited number of organizations.
Anthropic said it is currently investigating the matter and added that “there is no evidence that the access affected its core systems.”
The incident highlights the real-world risk of security vulnerabilities surrounding the control and proliferation of high-performance AI models.
Mythos was unveiled on April 7 and is being operated as part of Anthropic’s Project Glasswing.
The project is a restricted program that provides select companies with access to the unreleased model Claude Mythos Preview for defensive cybersecurity testing, such as identifying security vulnerabilities.
