Anthropic is expanding the defensive use of Claude Mythos 5, its most capable model for cybersecurity, by integrating it into Claude Security and preparing to bring the model’s capabilities into third-party cyber-defense products. The move is designed to give security teams access to frontier AI for finding software vulnerabilities and recommending fixes without providing unrestricted direct access to the powerful model.
Claude Security scans can now run on Mythos 5 for Claude Enterprise customers as part of a public beta. Alongside the integration, Anthropic is launching a $35 million Defender Advantage Fund for open-source security and plans to expand its Cyber Verification Program. Together, the initiatives reflect Anthropic’s attempt to make increasingly powerful AI useful to defenders while limiting the risk that the same capabilities could be repurposed for offensive cyber operations.
Claude Security Now Runs On Mythos 5
Claude Security is Anthropic’s vulnerability-scanning product for software development and security teams. Enterprise users can select a repository and have Mythos 5 analyze the codebase for security weaknesses.
The system returns structured findings rather than giving the user unrestricted access to Mythos 5. Each vulnerability report includes a Common Weakness Enumeration, or CWE, category, along with confidence and severity ratings and a suggested fix.
Users can subsequently open Claude Code on the web to implement a proposed fix. However, that interactive process uses models already available to the organization through Claude Code rather than extending direct Mythos 5 access.
Every patch also requires human review and approval before implementation.
Claude Security With Mythos 5 At A Glance
| Feature | Details |
|---|---|
| AI model | Claude Mythos 5 |
| Product | Claude Security |
| Availability | Public beta |
| Eligible customers | Claude Enterprise |
| Main task | Codebase vulnerability scanning |
| Findings include | CWE category, confidence and severity |
| Remediation | Suggested fixes |
| Patch approval | Human review required |
| Mythos direct access | Not provided through the scan |
| Billing | Standard token usage under existing Enterprise plan |
The structure is significant because Anthropic is separating access to the model’s capabilities from direct access to the model itself. Customers receive the defensive output they need while Mythos operates in the background.
Why Anthropic Is Restricting Direct Mythos 5 Access
Mythos 5 is unusually capable in areas that have substantial dual-use potential. Anthropic describes it as its most capable model for cybersecurity and biology research, with particularly strong capabilities in areas including exploit reasoning.
Those capabilities can help defenders discover vulnerabilities before attackers exploit them. But similar reasoning capabilities could potentially assist malicious users attempting to identify weaknesses or develop attacks.
As a result, Mythos 5 has not been released in the same unrestricted manner as a conventional general-purpose AI model.
Mythos 5 Vs Fable 5
| Metric | Claude Mythos 5 | Claude Fable 5 |
|---|---|---|
| Underlying model | Mythos-class model | Same underlying model |
| General availability | Limited | Broadly available |
| Cyber safeguards | Reduced for vetted use | Additional safety classifiers |
| Main positioning | Cybersecurity and specialized research | General reasoning and agentic work |
| Context window | 1 million tokens | 1 million tokens |
| Maximum output | Up to 128,000 tokens | Up to 128,000 tokens |
| Input price | $10 per million tokens | $10 per million tokens |
| Output price | $50 per million tokens | $50 per million tokens |
Fable 5 and Mythos 5 share the same core capabilities and technical specifications, according to Anthropic’s documentation. The major difference is the safety layer and availability model. Fable 5 incorporates additional classifiers designed to restrict risky requests, while Mythos 5 is provided through more controlled access arrangements.
From Project Glasswing To Broader Cyber Defense
Anthropic’s controlled deployment of advanced cybersecurity AI began before the latest Claude Security integration.
In April 2026, the company launched Project Glasswing to provide a small group of organizations responsible for critical software and infrastructure with access to Claude Mythos Preview and subsequently Mythos 5.
The objective was to give defenders an opportunity to discover and patch vulnerabilities before similarly capable AI systems became widely available or potentially accessible to attackers.
Anthropic’s Mythos Cyber-Defense Expansion
| Initiative | Purpose |
|---|---|
| Project Glasswing | Controlled Mythos access for critical defenders |
| Claude Security | AI vulnerability scanning for Enterprise customers |
| Cybersecurity partnerships | Embed Mythos into existing security products |
| Defender Advantage Fund | Support open-source security projects |
| Cyber Verification Program | Provide vetted defenders broader dual-use capabilities |
Anthropic says it is continuing Project Glasswing in collaboration with US government partners, focusing particularly on organizations protecting critically important infrastructure that meet strict security requirements.
Mozilla Experience Shows Potential Scale
Anthropic has already provided examples of how advanced Claude models can change vulnerability discovery.
The company’s cybersecurity materials say Mozilla shipped fixes for vulnerabilities discovered using Claude Opus 4.6 in March 2026. When Mythos Preview was subsequently used, Mozilla shipped another 271 fixes in its April release.
Anthropic says that figure was more than 20 times Mozilla’s normal monthly average.
Reported Mozilla Security Results
| Metric | Reported Result |
|---|---|
| Model used in later phase | Claude Mythos Preview |
| Additional fixes shipped | 271 |
| Comparison with normal monthly average | More than 20x |
| Deployment | Mozilla April release |
| Final patch decisions | Human-reviewed |
The example demonstrates the potential productivity benefit of AI-assisted security analysis: models can inspect large quantities of code and surface issues that human teams can then investigate and prioritize.
However, Anthropic’s approach continues to put humans in control of the final remediation decision.
$35 Million Fund Targets Open-Source Security
Anthropic is also committing substantial computing resources to securing open-source software.
The company’s new Defender Advantage Fund, abbreviated as 0xDAF, will provide $35 million in credits to organizations working on open-source security.
The fund will support projects using AI to identify and patch vulnerabilities, automate parts of open-source security workflows and experiment with new defensive techniques.
Defender Advantage Fund
| Fund Metric | Details |
|---|---|
| Name | Defender Advantage Fund |
| Abbreviation | 0xDAF |
| Commitment | $35 million |
| Support format | Credits |
| Primary focus | Open-source software security |
| Initial approach | Small number of larger pilot grants |
| Activities | Scanning, patching and security automation |
Open-source software represents an important target because widely used packages can sit deep within the software supply chains of thousands of organizations.
A vulnerability in one widely adopted dependency can potentially affect many downstream products. Improving automated vulnerability discovery could therefore produce security benefits beyond the organization maintaining the original project.
Mythos 5 Is Coming To Third-Party Security Tools
Anthropic does not intend to limit Mythos-powered security capabilities to Claude Security.
The company is working with cybersecurity technology and services partners to integrate Mythos 5 into products used for security operations, incident response, threat intelligence and detection engineering.
Instead of allowing users to directly interact with Mythos, these products can use the model in the background for specific tasks.
For example, a vulnerability-remediation product might return a list of suggested patches generated by Mythos without allowing the user to ask the underlying model to develop an exploit.
Anthropic’s Controlled-Access Model
| Access Method | User Receives | Direct Mythos Access |
|---|---|---|
| Claude Security scan | Vulnerability findings and suggested fixes | No |
| Partner security product | Task-specific security output | No |
| Project Glasswing | Controlled advanced capabilities | Limited to approved partners |
| Cyber Verification Program | Broader capabilities for vetted defenders | Controlled |
| General Claude users | Safeguarded models such as Fable 5 | No unrestricted Mythos access |
This architecture represents an important approach to managing dual-use AI. Instead of treating model access as an all-or-nothing decision, Anthropic is attempting to expose specific useful capabilities while keeping more dangerous functionality behind controlled interfaces.
Cyber Verification Program Will Expand
Anthropic is also expanding its Cyber Verification Program.
The existing program provides approved cybersecurity organizations with reduced safeguards when using Claude Opus and Sonnet models, allowing legitimate security professionals to perform dual-use tasks with fewer interruptions.
Over the coming weeks, Anthropic plans to expand defensive capabilities such as vulnerability triage and validation toward Mythos-class models, while broader Mythos access is expected to follow under controlled conditions.
The company says existing organizations accepted into the program do not need to reapply.
AI Could Reshape Vulnerability Discovery
The broader significance of Mythos 5 extends beyond Anthropic’s own products.
AI systems capable of analyzing enormous repositories, reasoning across complex software dependencies and proposing patches could dramatically increase the amount of code that security teams can examine.
Traditional security teams face an enormous asymmetry: software ecosystems contain millions of lines of code and dependencies, while defenders have limited time and personnel to inspect them.
Frontier models potentially change that equation by allowing organizations to continuously analyze much larger portions of their software estate.
At the same time, improved vulnerability discovery is inherently dual-use. A model capable of identifying previously unknown weaknesses for defenders could potentially provide similar advantages to attackers if unrestricted.
That tension explains why Anthropic is emphasizing controlled deployment rather than simply releasing Mythos 5 broadly.
The Bigger Picture
Anthropic’s Mythos 5 strategy illustrates an emerging challenge for frontier AI companies: the same capability improvements that make models more valuable for cybersecurity can also increase their potential for misuse. Mythos 5’s advanced exploit reasoning makes it particularly useful for finding weaknesses, but it also creates reasons to restrict unrestricted access.
The company’s response is to distribute outcomes rather than unrestricted capability. Claude Security customers can receive vulnerability findings and patches, partner products can use Mythos for narrowly defined defensive tasks, and vetted organizations can receive deeper access through programs such as Project Glasswing. The $35 million open-source fund extends the strategy by putting additional resources behind defensive applications.
Looking Ahead
The effectiveness of this approach will depend on whether Anthropic can expand defensive access without making Mythos 5’s more dangerous capabilities easier to misuse. Integrating the model into existing cybersecurity products could significantly increase its reach because security teams would not need to adopt an entirely new workflow or directly operate the underlying model. Anthropic says these integrations are still at an early stage and expects them to expand over time.
The larger race will be between AI-assisted attackers and AI-assisted defenders. As frontier models become better at understanding complex software, both vulnerability discovery and exploitation could accelerate. Anthropic’s bet is that controlled access to models such as Mythos 5 can give legitimate defenders enough of that capability to find and fix vulnerabilities earlier while keeping the most dangerous functionality away from malicious actors.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
