The Most Dangerous Network Weapon Model Mythos by Anthropic, Known for "Safety First," Suffers Supply Chain Breach

[Fact Check·Source: Google Verification, X Platform Public Signals] Recently, AI company Anthropic confirmed that their unreleased product Mythos, labeled as the "most dangerous model," has been hacked. Hackers gained access to the model and other unreleased models by guessing URL patterns and using credentials from third-party contractor Mercor. The hackers claim to have obtained full model pipeline access. The model is publicly described as capable of infiltrating mainstream operating systems and browsers as a network-level weapon. As of this report, Anthropic has only stated that they are investigating access reports from third-party vendor environments, without disclosing the specific capabilities of the Mythos model, the extent of data leakage, or subsequent security rectification measures. The actual impact of the incident remains highly uncertain [Fact Check·Source: Anthropic Public Statement].

The Incident Exposes Structural Misalignment in High-Risk AI Firms' Security Systems

The core reason the incident caused public outcry is that Anthropic has long branded itself with "Constitutional AI" and "Safety First" as its core brand labels, recognized as an industry benchmark for AI safety. The fact that such a firm could experience a high-risk model leak directly undermines public trust in the security capabilities of leading AI labs. Long-term tracking by winzheng.com reveals that such incidents are essentially the inevitable result of the industry's prevalent "heavy model alignment, light boundary protection" security investment misalignment.

According to the winzheng.com YZ Index v6 assessment, the main ranking capabilities of leading generative AI firms, such as code execution and material constraints, have an average score of 87.3 points. In the supply chain segment, only 62% of firms receive a pass on integrity ratings, with the remaining 38% marked as warn or fail. The stability of model output (i.e., the standard deviation of response results) averages 27%, with output consistency in high-risk scenarios below 70%. The engineering judgment (side ranking, AI-assisted assessment) score for high-risk model protection is only 59 points, indicating a pervasive issue of security investment tilting towards model alignment, with insufficient investment in supply chain and access control.

AI security research institution Apart Research's researcher publicly stated: "If even companies that put 'safety first' into their core strategies cannot protect their most dangerous models, the current voluntary AI governance framework is essentially ineffective."

winzheng.com has long advocated that the security protection of high-risk AI should not remain solely at the model alignment level but must cover the entire chain from "R&D-supply chain-storage-access." In this incident, the hackers neither broke Anthropic's core encryption system nor breached the model's security alignment; they merely exploited weak credentials from a third-party supplier to gain access to the most dangerous model's permissions, highlighting the necessity of full-chain protection.

[winzheng.com Independent Judgment] This incident is not a sporadic third-party security accident but a result of the inevitable mismatch between the rapid expansion of high-risk AI technology and the lag in security capabilities. We recommend that global regulatory bodies promptly issue mandatory regulations requiring AI firms developing high-risk capabilities such as cyberattacks and biosynthesis to meet three major requirements: first, access permissions for third-party suppliers must follow the "minimal necessary + dynamic zeroing" principle; second, storage of high-risk models must be physically isolated from networks; third, all high-risk AI R&D projects must submit a complete full-chain security protection plan to regulatory authorities. This incident also provides a crucial reference sample for China in formulating AI governance rules and improving the high-risk AI regulatory system.