Nuclear-grade vulnerability out of control? Anthropics' most powerful AI model "Mythos" exposed for unauthorized access, regulatory containment imminent.

According to reports, a small group of unauthorized users accessed the new Mythos AI model of Anthropic PBC company. The company had warned that this model is extremely powerful and could be used for carrying out high-risk cyber attacks. Sources revealed that on the same day Anthropic announced plans to release the model to a few companies for testing, a few users in a private network forum had already gained access to Mythos. Since then, this group of users has been regularly using Mythos, but not for cybersecurity purposes. Screenshots and demonstrations provided by the source confirmed this. Anthropic stated that Mythos can identify and exploit vulnerabilities in "all mainstream operating systems and browsers" at the user's direction. Therefore, the company intentionally restricted access to this technology through a project called the "Glasswing Project," ensuring that only specific software developers could test and strengthen their systems to prevent potential cyber attacks. This undisclosed leak incident highlights the serious challenges that Anthropic faces in controlling the spread of cutting-edge high-risk AI technology. This incident has also raised questions about whether there are other individuals using Mythos without authorization and for what purposes. Sources revealed that these users mainly accessed Mythos by using third-party contractor permissions, network information reconnaissance tools, and model address format information obtained from the Mercor data leak incident. These users belong to a private Discord group focused on discovering unreleased AI models. The sources also stated that these users are only researching new models out of interest and have not engaged in malicious activities, mainly conducting simple tasks such as building web pages to avoid detection. Additionally, these individuals have also gained access to several other unreleased models from Anthropic. A spokesperson for Anthropic stated, "We are investigating a report claiming that someone accessed Claude Mythos Preview without authorization through our third-party supplier environment." The company mentioned that there is currently no evidence to suggest that the access mentioned in the report exceeded the scope of the third-party supplier environment, and there is no evidence that this access affected any of Anthropic's systems. Currently, Anthropic has allowed Apple, Amazon, Cisco Systems, and dozens of other institutions to start testing Mythos. As a key partner and sponsor of Anthropic, Amazon also provides Mythos services to approved institutions through its Bedrock platform. Recently, financial institutions and government agencies in Europe and the United States are rushing to apply to join the Mythos testing list to enhance their own cybersecurity defenses. It is worth noting that the early disclosure of Mythos testing information caused market fluctuations: in late March, as the potential risks of the model being able to autonomously discover vulnerabilities and bypass existing network defenses gradually surfaced, market concerns grew, and the cybersecurity sector collectively came under pressure, with stocks like CrowdStrike, Palo Alto Networks, and Zscaler experiencing significant declines. Due to limited understanding of the model's actual functionality and inability to obtain access, banks and government agencies are also widely concerned. It is reported that US Treasury Secretary Yellen and Federal Reserve Chairman Powell urgently convened a closed-door meeting with executives from Wall Street giants such as Citi, Morgan Stanley, Bank of America, Goldman Sachs, and JPMorgan; meanwhile, Bank of England Governor Bailey called for global regulatory agencies to quickly assess the threat posed by the model, and the Bank of Canada has also convened discussions with financial institutions on the related risks.