Bringing the cybersecurity capabilities of Claude Mythos 5 to more defenders

We're sharing an update connected our efforts to thief much teams usage frontier capabilities for cyber defense. Claude Mythos 5 is now disposable successful Claude Security, and coming soon to partners' cyber defense tools. We're besides launching a $35M money to thief unafraid open-source package and sharing plans to grow our Cyber Verification Program.

In April, we launched Project Glasswing to put our astir tin frontier model, Claude Mythos Preview (and its successor, Claude Mythos 5), successful the hands of a mini group of organizations securing the world’s astir captious software. This gave defenders a model of clip to find and hole vulnerabilities up of models pinch akin capabilities becoming mostly disposable aliases reaching malicious actors.

Our extremity has ever been to grow Mythos-level defense to arsenic galore defenders arsenic we safely can. To do that, we've been moving connected safety classifiers and safeguards that fto america grow entree to Mythos-class models without putting their violative cyber capabilities successful the incorrect hands. Claude Fable 5 was the first step: it made the exemplary broadly disposable while blocking dual-use cyber work.

Today, we’re taking the adjacent steps. The riskiest behaviour occurs erstwhile a personification has nonstop entree to a model, wherever a malicious character tin effort to steer it toward harmful uses. But if users tin only person circumstantial outputs, specified arsenic a spot for a vulnerability aliases a information alert, that consequence is overmuch lower. The changes we’re announcing springiness users greater entree to the protect results, while maintaining due guardrails astir nonstop entree to the model:

• Claude Mythos 5 integration into the devices defenders trust on. We’re moving pinch our cybersecurity exertion and services partners to merge Claude Mythos 5 into the products and services defenders already usage to unafraid their software.
• Claude Security scans tin now tally connected Claude Mythos 5. Customers connected Claude Enterprise plans tin now tally our astir tin exemplary successful Claude Security, utilizing it to scan their codebases for information vulnerabilities and propose patches.
• $35 cardinal successful credits for open-source security. Our caller Defender Advantage Fund (0xDAF) will supply $35 cardinal successful credits to organizations moving to spot vulnerabilities successful open-source projects, automate parts of the process of scanning and patching open-source software, and research pinch caller information approaches. 
• Expanding our Cyber Verification Program. The programme already gives vetted defenders reduced safeguards connected Opus and Sonnet models. In the coming weeks, we will grow this programme to see broader dual-use capabilities connected Opus and Sonnet, pinch Mythos-class entree to follow. 

Our purpose remains to thief organizations accommodate to the gait and demands of cybersecurity arsenic AI models go progressively powerful. We will proceed to create safeguards, entree programs, and organization support to make our astir tin models safely disposable to a wide scope of group and organizations.

Integrating Mythos into existing cyberdefensive tools

The teams defending hospitals, utilities, financial systems, and the package proviso concatenation already trust connected a suite of products and services for information operations, incident response, threat intelligence, and discovery engineering. The fastest measurement to make frontier capabilities disposable to those defenders is to merge Mythos-class models into the devices they already run.

Many of our partners person already built cyber products connected Claude Opus that thief information teams triage alerts, place threats, and remediate vulnerabilities faster. We’re now moving pinch these partners and much to build Claude Mythos 5 into their products and services, truthful they tin present Mythos-level protect outcomes to their customers. 

When an extremity personification uses 1 of these products, they’re not interacting pinch Mythos directly. Instead, they activity done a purpose-built interface that runs Mythos successful the inheritance for a defined task and only person the circumstantial artifact the merchandise is intended to provide. For example, a instrumentality to remediate vulnerabilities mightiness supply a database of suggested patches arsenic its output. This output would beryllium generated by Mythos, but the personification would not person a measurement to punctual the exemplary to, say, create an utilization for a vulnerability. We and our partners besides person maltreatment prevention measures successful spot to verify the exemplary stays wrong its intended scope.

We're early successful this activity and expect it to grow complete time. If you build information products aliases services and want to bring Claude Mythos 5 to your customers, you tin register your liking here. 

Making Claude Security disposable pinch Claude Mythos 5 for Enterprise customers

Starting today, Claude Security scans now tally connected Claude Mythos 5. Claude Security scans codebases for vulnerabilities and suggests patches for quality review; it’s presently successful nationalist beta for Claude Enterprise customers, and scans pinch Mythos 5 are billed arsenic modular token usage nether your existing plan, pinch nary abstracted add-on. 

Enterprise admins tin alteration Claude Security successful the admin console. From claude.ai/security, users tin prime a repository to scan utilizing Claude Mythos 5. Claude past scans the codebase for vulnerabilities, and returns each uncovering pinch a CWE (Common Weakness Enumeration) category, assurance and severity ratings, and a suggested fix. 

Users tin past unfastened Claude Code connected the web to instrumentality the fix. Interactive patching uses the models your statement has entree to successful Claude Code. The Mythos scan itself does not widen Mythos entree to different surfaces. Every spot must beryllium reviewed and approved by a quality earlier it tin beryllium implemented.

Claude Security uses Mythos 5 to scan codification you own, and returns elaborate findings alternatively than earthy outputs without exposing the exemplary itself. This intends defenders tin entree the capabilities of Claude Mythos 5 without the exemplary becoming accessible to those who mightiness misuse it. 

For much astir Claude Security, spot our guide to getting started. 

Launching the Defender Advantage Fund to unafraid open-source software

Some of the world’s astir wide utilized programs tally connected open-source software. Yet these projects are often maintained by volunteers aliases nonprofit foundations, who whitethorn deficiency the resources aliases unit to comprehensively take sides their projects against attack. Through Project Glasswing, we made $4M successful nonstop donations to open-source information organizations, provided credits to the open-source information foundations successful the program, helped scan and spot wide utilized projects, and support coordinated vulnerability-fixing efforts for illustration Akrites and Gold Eagle.

Our caller Defender Advantage Fund (0xDAF) builds connected that activity pinch $35 cardinal successful Claude credits for organizations helping open-source maintainers unafraid their software. Grants will attraction connected 3 areas: patching unrecorded vulnerabilities successful wide utilized projects, automating scanning and patching successful ways different projects tin replicate, and helping projects prosecute much eager information approaches that make them resistant to full classes of attack.

We're starting pinch a mini number of larger, aviator grants to study what useful and scales best. We will stock specifications connected first recipients successful the coming weeks.

Expanding our Cyber Verification Program

To date, our Cyber Verification Program has provided organizations pinch entree to dual-use capabilities erstwhile utilizing Claude Opus and Sonnet models. Organizations successful the programme acquisition reduced safeguards, minimizing interruptions for accepted teams doing morganatic cybersecurity activity connected systems they’re authorized to protect.

Over the coming weeks, we are evolving the programme to grow safeguarded entree to Claude Mythos. As portion of this, entree to protect capabilities for illustration vulnerability triaging and validation will grow to Mythos-class models, and cyber defenders will spot reduced blocks connected Claude Opus and Sonnet-class models. Additionally, we are continuing to grow entree to Claude Mythos done Project Glasswing successful collaboration pinch our partners successful the U.S. Government, focused connected protectors of critically important infrastructure that meet strict information power requirements.

We'll stock much specifications astir the Cyber Verification Program description successful the coming weeks. In the meantime, we promote each information teams performing morganatic cybersecurity activity to use for the programme for reduced safeguards connected Claude Opus and Sonnet models. If you are already enrolled and accepted, nary action is needed; we’ll scope retired pinch updates.

What’s next

These initiatives are a continuation of our efforts to make the protect capabilities of frontier models disposable to much group and organizations, and to support the open-source organization successful hardening their projects against attack. We will proceed to activity pinch authorities partners, organizations, open-source maintainers, and the broader manufacture to build the resilient cyber infrastructure today’s highly tin AI models demand. 

• Apply for the Cyber Verification Program.
• Register your interest successful building cyber products and offerings pinch Mythos.
• Claude Security is disposable successful nationalist beta for Enterprise customers. Admins tin alteration Claude Security successful the admin console. For a afloat walkthrough, spot our guide to getting started.