Anthropic has crossed a significant threshold in cybersecurity with the launch of Claude Mythos Preview, an advanced, non-commercial vulnerability detection model. It has uncovered thousands of critical zero-day vulnerabilities across major operating systems and web browsers, many of which had eluded automated tools for two to three decades. In conjunction with this release, Anthropic is forming an unprecedented industry consortium and investing $100 million to operationalize these capabilities.
The ability of large language models to analyze and exploit code has been steadily improving over recent quarters, showing useful results alongside existing security tools without disrupting the fundamental balance between attackers and defenders. Mythos Preview represents a transformative shift; the model now operates entirely autonomously, requiring no human guidance, and produces results that even the best offensive security experts could not achieve faster.
The demonstrated capabilities released by Anthropic today are unprecedented in their scope. Within just a few weeks of testing, Mythos Preview identified a 27-year-old vulnerability in OpenBSD—a system renowned for its hardened security—allowing a remote attacker to shut down any exposed machine with a simple connection. It also uncovered a 16-year-old flaw in FFmpeg, a video processing library embedded in countless applications, which had evaded detection despite 5 million automated tests. On the Linux kernel, the model autonomously chained several vulnerabilities to enable privilege escalation, granting complete system control. All identified vulnerabilities have been reported and patched by the relevant maintainers ahead of the public announcement.
An Innovative Consortium to Support the Launch
An appropriate response to this breakthrough is strategically collective. Founding partners of Project Glasswing include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks. More than 40 organizations managing critical infrastructure or major open-source components will also gain early access to Mythos Preview for auditing their codebases. Anthropic’s financial commitment covers $100 million in usage credits, alongside $2.5 million allocated to Alpha-Omega and OpenSSF via the Linux Foundation, and $1.5 million to the Apache Software Foundation—organizations dedicated to securing the open-source components that form the backbone of modern software systems.
The strategy aims to maximize momentum and coverage for Mythos Preview by rallying global actors and investing a capital infusion of $100 million. Demand for cybersecurity will drive the initiative forward, as noted by Lee Klarich, CTO of Palo Alto Networks: “We will see more attacks, faster attacks, and more sophisticated attacks.”
Reducing the Gap Between Discovery and Exploitation
The cybersecurity capabilities of Mythos Preview stem from its advanced software engineering proficiency. On the SWE-bench Verified, a benchmark for the autonomous resolution of GitHub tickets, the model scored 93.9% compared to 80.8% for Claude Opus 4.6. On CyberGym, a benchmark focused on replicating vulnerabilities, it achieved 83.1% versus 66.6% for Opus 4.6. On GPQA Diamond, an expert reasoning model, its score rose to 94.6%. These margins are significant and indicate a qualitative leap in the model’s capability to reason about code in real contexts, beyond mere generation.
Microsoft has reported that Mythos Preview exhibits substantial improvements on CTI-REALM, its open-source security benchmark, compared to earlier generations. AWS indicates that it has already applied the model to critical internal codebases. CrowdStrike emphasizes that the time between the discovery of a vulnerability and its exploitation, which once spanned months, can now be reduced to a matter of minutes with the assistance of AI.
Restricted Access and Pre-Costing
Anthropic does not intend to make Mythos Preview available to the general public. The priority is to develop safeguards that can detect and block the model’s most dangerous outputs, a task that will run parallel to the development of a forthcoming model, Claude Opus, which carries lower residual risk. Following the research phase covered by the committed credits, access to Mythos Preview will be charged at $25 per million tokens for input and $125 for output, via the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.
For European organizations subject to NIS2 or DORA, a concern remains: the access channels for Mythos Preview—including the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry—are all operated by entities subject to the U.S. Cloud Act. Submitting critical infrastructure code for auditing by this model effectively means routing it through environments without enforceable guarantees of confidentiality under European law. Anthropic has not announced any access modalities certified by ANSSI or compliant with SecNumCloud requirements. This limitation is something that legal and security departments must address prior to any deployment.



