Anthropic Unveiled Mythos: AI for Vulnerability Discovery Is Already Reshaping Cybersecurity and Warfare

Mythos for Anthropic is not just another version of a chatbot, but a signal that AI has crossed a threshold where it can significantly change both cybersecurity and military planning. The model can not only find errors in code but also turn them into working exploits with minimal human involvement. This is precisely why the company did not give it mass public access, and Gregory Allen, senior adviser at the Wadhwani AI center at CSIS, directly states: such systems simultaneously strengthen defenses and make the world more dangerous in the transition period, while vulnerable software has not yet been rewritten and patched.

Anthropic announced Claude Mythos Preview on April 7, 2026. The company describes the model as an advanced general-purpose system with strong agent capabilities in coding and reasoning, but Mythos's key distinction is its level of cybercapabilities. According to Anthropic, the model has already found thousands of high-criticality vulnerabilities, including issues in every major operating system and major web browser.

In official materials, the company states that Mythos is capable of autonomously detecting and exploiting zero-day vulnerabilities. In internal and external testing, this looks like a qualitative leap: on the CyberGym benchmark, the model showed 83.1% versus 66.

6% for Anthropic's previous nearest model. Among already disclosed examples are a 27-year-old error in OpenBSD, a 16-year-old vulnerability in FFmpeg, and a chain of bugs in the Linux kernel. Due to dual-use concerns, Mythos was not released publicly.

Instead, Anthropic launched Project Glasswing—a limited program through which the model is provided to companies and organizations responsible for critical software and infrastructure. Among the initial partners, the company lists AWS, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, Palo Alto Networks, and the Linux Foundation. Additionally, access was given to more than 40 other organizations that maintain or build critical software.

Anthropic is allocating up to $100 million in Mythos usage credits for this program and another $4 million in direct donations to the open source ecosystem. The logic is straightforward: if such models soon become more widely available, defenders need at least a small time window to manage to close the most dangerous holes. Allen believes this time window is precisely the key story.

According to him, cybersecurity has lived for many years under a shortage of specialists, and good attackers and vulnerability researchers are too expensive and rare. If AI begins to perform a significant portion of this work itself, the entire economics of the industry changes. In an ideal scenario, almost every new software product will undergo something like a Mythos test before release: the model will attempt to break it before malicious actors do.

But between the current state of the internet and this scenario lies a vast layer of legacy code, including open source projects that rely on small teams and volunteers. Allen estimates the next 12 months as a period of large-scale and painful restructuring, especially for banks, energy companies, and other critical infrastructure owners. He also argues that the U.

S. federal government is far from creating a comparable advanced model and thus depends on private companies; in the case of Anthropic, according to his assessment, the gap from competitors could be anywhere from six to 18 months. The military side of the conversation sounds no less harsh.

Allen says that Anthropic models are useful not only in cyber defense but also in intelligence and combat operations, and claims that such AI capabilities are already being used by the U.S. in the war with Iran.

As a framework, he recalls Project Maven: initially this was an attempt to automate the primary analysis of video streams from drones and satellites to relieve analysts. Now, according to him, the combination of computer vision and large language models allows not just marking an object on screen but comparing changes over time, identifying anomalies, formulating a first version of an intelligence summary, and accelerating the entire decision-making cycle. Allen provides a striking figure: whereas previously the benchmark for the American system was dozens of strikes per day, in the first 24 hours of the Iranian campaign, he says, it was about 1,000 targets hit per 24 hours.

And behind each such figure stands not only the strikes themselves but thousands of decisions on processing intelligence data, identifying and prioritizing mobile and concealed targets. The main conclusion here is that Mythos is important not as a rare closed model in itself, but as a sign of a new phase. AI is beginning to replace scarce expert labor in the most sensitive areas—from code audits to intelligence analysis.

For the market, this means increased demand for accelerated vulnerability remediation, new secure-by-default development standards, and closer ties between AI labs, clouds, banks, and the state. For governments—an uncomfortable dependence on private developers who move faster than bureaucracy. And for everyone else—a short window when the same tools can still be used primarily for defense rather than attack.