Anthropic officially announced today (August 21, Eastern Time) that its most powerful cybersecurity model, Claude Mythos 5, has been integrated into the Claude Security service and is now available for public beta to all Claude Enterprise customers. At the same time, the company launched the Defender Advantage Fund (ticker: 0xDAF) with a total of $35 million (approximately NT$1.14 billion) to fund open-source software security maintenance efforts, and announced the expansion of the Cyber Verification Program’s scope.
Claude Mythos 5: From Closed to Enterprise Tool
Claude Mythos 5 is Anthropic’s most powerful model in the cybersecurity domain. Since April this year, it has been available only to a small number of vetted organizations through the Project Glasswing initiative, including U.S. government partners and critical infrastructure protectors. The model substantially surpasses general large language models in security capabilities such as vulnerability exploitation reasoning. However, due to its dual-use nature—applicable to both defense and offense—Anthropic has maintained strict access controls to prevent the model from falling into the hands of malicious actors.
According to Anthropic Official explanationThe most dangerous behaviors occur when users can interact directly with a model, as malicious users may attempt to steer the model toward generating harm. However, if users can only receive specific outputs—such as vulnerability patches or security alerts—the risk drops significantly. The key design choice of this integration is to give enterprise users access to Mythos 5’s analysis results through the purpose-built Claude Security interface, rather than granting direct access to the model itself. The recently released Claude Fable 5 already expanded model usage by blocking dual-use cybersecurity work, and the Claude Security integration takes that a step further.
In terms of the specific workflow, the administrator first enables the feature in the backend, and the team selects a GitHub repository from claude.ai/security to scan. Mythos 5 scans the code for vulnerabilities and returns the CWE (Common Weakness Enumeration) classification, confidence and severity ratings, and suggested patches for each finding. Users can then open the patches in Claude Code’s web version to implement them, but each patch must go through human review and approval before it can be applied. Anthropic emphasizes that interactive patching uses the model permissions the organization already has in Claude Code, and the Mythos scan itself does not extend model access to other conversations. Scan costs are calculated based on standard token usage under the existing Enterprise plan, with no additional purchase required.
Mythos Scan itself does not extend model access to other features; users cannot use Mythos 5 in general conversations simply because they have Security permissions. Anthropic’s statement on this is: “Defenders can gain Mythos 5’s capabilities, but the model itself will not be made available to those who might abuse it.” This is fully consistent with the company’s longstanding position: packaging the most powerful models’ capabilities into tools for enterprises, rather than directly opening the model for anyone to access.
US$35 million Defender Advantage Fund
Anthropic also announced the establishment of the Defender Advantage Fund (codename 0xDAF), with an initial commitment of $35 million in Claude credits to support open source software security maintenance. The fund will focus on three main areas: first, patching immediate vulnerabilities in widely used projects; second, automating scanning and patching workflows so other projects can replicate them; and third, helping projects adopt more advanced security strategies to fundamentally resist entire classes of attacks.
This fund is an extension of Project Glasswing. That initiative previously provided $4 million in direct grants and Claude credits to open-source security organizations, helping to scan and patch multiple widely used open-source projects, and supported community-led coordinated vulnerability remediation efforts such as Akrites and Gold Eagle. Anthropic said it will start with a small number of large pilot grants, learn which approaches are most effective and scalable, and announce the specific list of first grantee organizations in the coming weeks.
Partner Integration and Verification Program Expansion
At the same time, Anthropic will collaborate with multiple cybersecurity technology and service partners to integrate Mythos 5 into existing security products and services. Many partners have already built security tools based on Claude Opus, including alert classification, threat identification, and vulnerability remediation, and are now upgrading them to the Mythos tier. In this model, end users likewise will not interact directly with Mythos, but instead receive specific outputs through product-specific interfaces. Partners also have abuse prevention measures in place to ensure the model operates within expected parameters.
The Cyber Verification Program will also expand in the coming weeks. The program already provides verified organizations with reduced safety restrictions on the Opus and Sonnet models, minimizing friction during legitimate cybersecurity work. The new version will extend dual-use capability access to the Mythos-tier models and continue collaborating with the U.S. government through Project Glasswing, with a focus on critical infrastructure protectors. Organizations already in the Cyber Verification Program do not need to reapply—Anthropic will reach out proactively.
Anthropic also mentioned that they are continuously developing safety classifiers and safeguards with partners. These technologies are the foundation for expanding access to Mythos-level models, ensuring that defensive capabilities can spread to more people while preventing the models’ offensive capabilities from being misused. The company emphasized that their goal is to help organizations adapt to the cybersecurity pace and demands that come with increasingly powerful AI models, and they will continue to develop safeguards, access programs, and community support going forward.