Anthropic recently introduced a preliminary version of its advanced AI model, Mythos, which a select group of partner organizations will use for cybersecurity applications. An internal memo that was previously leaked described this AI startup’s model as one of its most potent creations to date.
This restricted launch is part of Project Glasswing, a new security initiative involving over 40 partner entities. These organizations will implement Mythos for “defensive security work” and to bolster the security of crucial software systems, as stated by Anthropic. Although Mythos was not explicitly trained for cybersecurity, this preview will be utilized to detect code vulnerabilities in both proprietary and open-source software, the company explained.
Anthropic asserts that Mythos has uncovered “thousands of zero-day vulnerabilities, many of them critical” over the last few weeks. The company noted that many of these vulnerabilities have existed for one to two decades.
Mythos functions as a general-purpose model within Anthropic’s Claude AI ecosystem, boasting strong capabilities in agentic coding and reasoning, according to the company. Anthropic’s frontier models are considered its most advanced and high-performing, engineered for intricate tasks such as agent development and coding.
The organizations participating in the Mythos preview include major players like Amazon, Apple, Broadcom, Cisco, CrowdStrike, the Linux Foundation, Microsoft, and Palo Alto Networks. Through this initiative, these partners will eventually share their insights from using the model, aiming to benefit the broader tech industry. Anthropic clarified that this preview will not be made generally available.
Furthermore, Anthropic claims to have engaged in “ongoing discussions” with federal officials regarding Mythos’s deployment. However, these conversations are likely complicated by the ongoing legal dispute between Anthropic and the Trump administration. The Pentagon designated the AI lab a supply-chain risk after Anthropic declined to permit autonomous targeting or surveillance of U.S. citizens, a refusal that led to retaliation.
The existence of Mythos was initially revealed through a data security incident reported by Fortune last month. A draft blog post about the model, then named “Capybara,” was inadvertently left in an exposed cache of documents within a publicly accessible data lake. This leak, which Anthropic later attributed to “human error,” was first detected by security researchers. According to the report, the leaked document stated, “‘Capybara’ is a new name for a new tier of model: larger and more intelligent than our Opus models — which were, until now, our most powerful,” subsequently adding that it was “by far the most powerful AI model we’ve ever developed.”
The leaked information indicated that Anthropic’s new model significantly surpassed the performance of its existing public models in areas such as “software coding, academic reasoning, and cybersecurity.” It also highlighted the potential cybersecurity risk if malicious actors were to weaponize the model to discover and exploit bugs, rather than rectify them—the intended use for Mythos.
In the preceding month, the company accidentally 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. Following this, the company then inadvertently caused thousands of GitHub repositories to be taken offline while attempting to resolve the initial mishap.