Anthropic unveiled a preview of its latest frontier model, Mythos, on Tuesday, announcing its initial deployment with a select group of partner organizations for critical cybersecurity operations. A previously leaked internal memo from the AI startup characterized Mythos as one of its "most powerful" models to date.
This restricted launch is an integral component of Project Glasswing, a new security initiative. Under this project, over 40 partner organizations will integrate Mythos to conduct "defensive security work" and bolster the security of essential software. Anthropic clarified that while Mythos was not explicitly trained for cybersecurity tasks, this preview version will be utilized to scan both proprietary and open-source software systems for potential code vulnerabilities.
Anthropic asserts that in recent weeks, Mythos has successfully identified "thousands of zero-day vulnerabilities, many of them critical." The company further noted that a significant number of these vulnerabilities date back one to two decades.
Mythos functions as a general-purpose model within Anthropic’s Claude AI ecosystem, boasting what the company describes as robust agentic coding and reasoning capabilities. Anthropic's frontier models are recognized as its most sophisticated and high-performance offerings, specifically engineered for tackling complex assignments, including agent construction and advanced coding.
The distinguished roster of partner organizations participating in the Mythos preview includes industry giants such as Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, and Palo Alto Networks. A key aspect of this initiative is that these partners will ultimately share their insights and learnings derived from using the model, aiming to benefit the broader tech industry. Anthropic confirmed that this preview will not be made generally available to the public.
Anthropic also stated that it has been engaged in "ongoing discussions" with federal officials regarding the applications of Mythos. However, these discussions are presumably complicated by the ongoing legal dispute between Anthropic and the Trump administration. This conflict arose after the Pentagon classified the AI lab as a supply-chain risk due to Anthropic’s refusal to permit autonomous targeting or surveillance of U.S. citizens.
The existence of Mythos was initially revealed last month through a data security incident reported by Fortune. A draft blog post concerning the model, then referred to as "Capybara," was discovered in an unsecured cache of documents accessible on a publicly inspectable data lake. Anthropic subsequently attributed this leak, first identified by security researchers, to "human error." The leaked document described "‘Capybara’ as a new name for a new tier of model: larger and more intelligent than our Opus models — which were, until now, our most powerful," further asserting it was "by far the most powerful AI model we’ve ever developed," according to the report.
Within the leaked information, Anthropic claimed its new model significantly surpassed the performance benchmarks of its currently public models in areas such as "software coding, academic reasoning, and cybersecurity." The document also highlighted a potential cybersecurity risk if the model were to be weaponized by malicious actors to discover and exploit bugs, rather than to fix them, which is the intended deployment of Mythos.
These revelations follow a series of recent incidents for the company. Last month, Anthropic inadvertently exposed nearly 2,000 source code files and over half a million lines of code due to an error during the launch of version 2.1.88 of its Claude Code software package. Subsequently, in its attempts to rectify this oversight, the company accidentally caused thousands of code repositories on GitHub to be taken offline.
The Editorial Staff at AIChief is a team of professional content writers with extensive experience in AI and marketing. Founded in 2025, AIChief has quickly grown into the largest free AI resource hub in the industry.