What is Claude Mythos Preview?

• Claude Mythos Preview, the official name of the Anthropic cybersecurity model, is a frontier LLM capable of finding and exploiting critical software flaws at machine speed. It has uncovered subtle, decades-old zero-day vulnerabilities that have eluded manual and automated scans over the years.
• With a reported 83% success rate in weaving together minor, low-severity vulnerabilities into critical control-flow hacks, Claude Mythos Preview collapses the traditional “time-to-exploit” window from weeks or months down to mere minutes.
• To give cyber defenders a head start on using Claude Mythos Preview and to prevent it from falling into the wrong hands, Anthropic launched Project Glasswing, which restricts access to a vetted defensive coalition of technology vendors, industry groups, and government agencies.
• In the Claude Mythos Preview cybersecurity era, increasingly advanced agentic cybersecurity capabilities will become publicly available and mainstream. Thus, cybersecurity teams must abandon reactive patching and adopt an AI-driven exposure management model, relying on continuous asset visibility, advanced attack path analysis, and automated, agentic remediation to harden networks before automated threats strike.

Anthropic’s Claude Mythos Preview is a frontier AI large language model (LLM) whose capabilities include execution of massive and complex tasks, as well as advanced reasoning, such as long-term planning, problem anticipation, and strategy iteration. 

As a general-purpose LLM, its scope of knowledge and expertise spans multiple areas, but what makes Claude Mythos Preview significant are its highly advanced cybersecurity capabilities: specifically, its ability to find and exploit software vulnerabilities at unprecedented speed.

When Anthropic announced Claude Mythos Preview on April 7, 2026, the AI company said this new AI model had found thousands of years-old security flaws that millions of previous manual and automated tests had missed. The security issues included misconfigurations, logic flaws, architecture gaps, and high-severity zero-day vulnerabilities impacting every major operating system and web browser. The oldest one it discovered was a 27-year old vulnerability in the widely used OpenBSD operating system.

Anthropic also announced that Claude Mythos Preview can create sophisticated exploits, even when prompted by users without formal cybersecurity training. Specifically, Claude Mythos Preview boasts an 83% autonomous success rate in chaining minor security flaws into critical, working exploits. Some researchers, however, have expressed skepticism about this statistic, which comes from Anthropic's internal testing and hasn’t been independently verified.

Consequently, Claude Mythos Preview has become a major topic in cybersecurity news coverage and among corporate boards of directors. The attention on Claude Mythos Preview has extended far beyond the technology industry.

While Claude Mythos Preview is often described as an agentic AI tool that can act autonomously, Tenable has found in its own testing of Mythos Preview and other frontier models that AI tools labeled as autonomous still require a considerable amount of human intervention. No LLM today is able to act with complete autonomy, although that is the future goal of LLM developers.

Anthropic describes Claude Fable 5 as a “Mythos-class” model that is “safe for general use” because it has more usage guardrails and controls that rein in its output. Anthropic released Claude Fable 5 in early June 2026.

Frontier models like Claude Mythos Preview are widely expected to cause an spike in vulnerability discovery. Years before the arrival of frontier AI models, the annual number of vulnerability disclosures often had made it difficult for organizations to patch all or most of the Common Vulnerabilities and Exposures (CVEs) in their environment. 

For example, about 50,000 new vulnerabilities were discovered in 2025, a 25% increase from 2024’s 40,000, already a staggering number. The Forum of Incident Response and Security Teams (FIRST), which operates the Common Vulnerability Scoring System (CVSS), forecasts a whopping 66,000 CVE discoveries in 2026. 

“As we look toward the second half of 2026, the vulnerability coordination domain is undergoing an unprecedented transformation. With the recent deployments of highly autonomous AI discovery tools, such as Anthropic’s Mythos […] and OpenAI’s GPT-5.4-Cyber, the volume of identified software flaws has accelerated massively,” FIRST wrote in its blog “The 2026 Vulnerability Forecast Update: Navigating the AI Epoch.”

However, it bears pointing out that just about 1% of vulnerabilities were confirmed as exploited in the wild in 2025, according to VulnCheck’s Exploit Intelligence Report.

As happens with almost everything touched by AI, the volume and pace of CVE discoveries have changed forever, and cybersecurity teams need to adjust to this new reality as soon as possible.

While cybersecurity teams have been prioritizing the remediation of the riskiest vulnerabilities for their organizations for years, frontier AI models like Claude Mythos Preview further reinforce the need to pinpoint the CVEs whose exploitation would cause the biggest disruption to their organizations.

Claude Mythos Preview’s ability to independently analyze large software codebases, discover hidden vulnerabilities, and generate sophisticated exploits in just hours has upended the traditional vulnerability management lifecycle by collapsing the time defenders have to react.

The time-to-exploit window — the time defenders have to react between the discovery of a security flaw and its active weaponization in the wild — has irreversibly and drastically shrunk from weeks or months to minutes, making traditional timelines for patching and remediation obsolete.

Managing and prioritizing vulnerabilities in a landscape reshaped by Claude Mythos Preview and other frontier LLMs requires a shift in three fundamental areas:

1. Surviving the “vulnerability tsunami”

The sheer volume of newly reported CVEs entering the pipeline at machine speed thanks to AI discovery risks overwhelming legacy enterprise workflows that rely on manual human triage and treat vulnerability management as a reactive ticketing exercise. Cybersecurity teams must treat vulnerability management as one component of continuous exposure management that also includes other risk areas like misconfigurations and identity flaws, and that maps the interconnections between them.

2. Moving from isolated severity scores to context-rich prioritization

Claude Mythos Preview excels at analyzing complex infrastructure webs holistically to identify how multiple benign-looking security issues can be stitched together to trigger a catastrophic breach. Because of this, the evaluation of Asset Criticality Rating (ACR), Asset Exposure Score (AES) and Vulnerability Priority Rating (VPR) vs. CVSS is no longer a theoretical debate: relying on static CVSS severity scores to prioritize patches is a failing strategy. Instead, cyber defenders must use context-rich vulnerability prioritization that calculates risk based on:

• Real-world threat intelligence
• Actual asset exploitability
• The asset's unique business context

3. Beating the AI clock with agentic remediation

To survive in the Claude Mythos Preview era, organizations must shift from reactive patching toward automated and autonomous remediation that matches high-priority CVEs with validated updates, coordinates complex multi-step workflows, and deploys custom fixes in real time.

When an adversarial frontier model can discover a zero-day vulnerability, assemble an exploit chain, and launch an attack in an automated loop, conventional patch management manually led by humans is simply too slow. To keep up, organizations should shift toward automated patch management, vulnerability remediation, and agentic AI remediation.

By deploying autonomous security tools — such as the Tenable Hexa AI custom agents that are part of the Tenable One Exposure Management Platform — defenders can securely connect LLMs directly to their tech stacks. This allows defensive AI to automatically and autonomously match high-priority CVEs with validated patches, synchronize complex multi-step workflows, and deploy custom fixes in real time.

The bottom line: Frontier LLMs like Claude Mythos Preview make it important for organizations to transition from a reactive patching posture to a preemptive cybersecurity model — using defensive AI to continuously assess, prioritize, and harden the attack surface before an offensive frontier LLM can target their environment.

With an AI-driven exposure management program that offers agentic AI capabilities, you can automate and accelerate detection, prioritization, and remediation of vulnerabilities to keep pace with the discovery speed of frontier LLMs like Claude Mythos Preview.

Frontier LLMs like Claude Mythos Preview introduce an entirely new class of cyber risks. When managing a frontier AI model with advanced, multi-step reasoning capabilities and the ability to act autonomously, traditional security protocols fall short. While AI vendors like Anthropic and OpenAI put security guardrails on their LLMs, organizations still must deploy specific AI governance policies and preemptive cybersecurity risk management practices to mitigate these unique threats.

How much cyber risk does AI create for organizations? Over a 30 day period, Tenable detected 457 million AI-related security issues among 7,000-plus organizations, an average of 62,000 exposures per organization.

Frontier LLM security presents significant challenges, including:

Unmanaged agentic autonomy

Unlike early AI chatbots that merely processed prompts and compiled answers, a frontier model possesses specialized, autonomous reasoning. When given a high-level goal, agentic AI tools can independently execute multi-step attack chains, write scripts, and interact with infrastructure without human intervention. This creates a severe visibility gap if the model operates outside the scope of traditional IT asset monitoring.

Automated exploit chaining and much higher number of CVE discoveries

AI is significantly boosting vulnerability discovery. Given this reality, attackers can use LLMs to continuously scan an enterprise’s attack surface and chain together low-severity software vulnerabilities, cloud misconfigurations, and identity weaknesses into a catastrophic critical system hijack. In short, they can abuse and manipulate frontier LLMs into crafting customized, novel attack paths for which cyber defenders have no existing mitigation plans.

The zero-day and “shadow AI” risk combination

Enterprises are eagerly adopting AI tools to streamline, automate, and accelerate all types of processes, a trend that involves the use not only of approved AI systems but also the use of unapproved AI tools, also known as “shadow AI.” Cybersecurity teams are blind to unmapped, unmanaged “shadow AI” deployments. Because they’re unprotected, shadow AI assets create a rapidly expanding, invisible attack surface.

In turn, frontier AI models like Claude Mythos Preview can autonomously collapse the window between a zero day vulnerability discovery and its active weaponization down to hours or minutes, making any unprotected AI pipeline or model repository an open gateway into the corporate network.

Given these unique cyber risks associated with the use of frontier AI models, organizations must shift to an operating model centered on continuous, AI-driven exposure management that allows them to preemptively remediate vulnerabilities, misconfigurations, and other security issues before attackers can exploit them.

Project Glasswing is an Anthropic initiative designed to grant a select group of technology industry and government leaders access to Claude Mythos Preview, while restricting everyone else’s access to the frontier AI LLM.

With an initial roster of 50 partners, Project Glasswing now has about 200 partners in about 15 countries. As of May 2026, Project Glasswing participants have used Claude Mythos Preview to find more than 10,000 high- or critical-severity security flaws in their codebases. They’re also using Claude Mythos Preview to:

• Create vulnerability patches
• Check for vulnerabilities before they ship code to production
• Conduct penetration testing on their infrastructure
• Automate threat detection and response
• Revamp legacy codebases using modern, memory-safe languages

Project Glasswing participants also get the opportunity to study how this frontier AI model interacts with code; to evaluate its behaviors; and to build defensive controls.

Before announcing Claude Mythos Preview, Anthropic recognized that the frontier AI model’s rare cybersecurity prowess could allow bad actors to find and exploit zero-day vulnerabilities in bulk and with unprecedented speed. 

That’s why via Project Glasswing, Anthropic has limited Claude Mythos’ access to handpicked organizations, including technology vendors, such as Tenable; technology industry groups, such as the Linux Foundation; and government agencies. 

Project Glasswing mirrors similar restrictive defensive programs in the industry, such as OpenAI TAC (Trusted Access for Cyber), which tightly controls access to models like GPT-5.4-Cyber for verified defenders.

Is Tenable a member of Project Glasswing? What is Tenable’s role? 

Yes, Tenable is a member of Project Glasswing. By participating in Project Glasswing and working with Claude Mythos Preview, Tenable can help customers better understand how emerging frontier AI models behave, their evolving risks and benefits for cybersecurity, and the kinds of controls organizations will need as AI adoption accelerates.

As part of Project Glasswing, we’re particularly interested in driving new research using Mythos Preview to better understand where it can help reinforce existing security analysis, and strengthen our own defenses by using frontier models to improve the security of Tenable. We also plan to use Mythos alongside other models to help challenge assumptions, and identify relationships and risk patterns faster than traditional approaches alone.

Tenable views its Project Glasswing participation, as well as Tenable’s ongoing collaborations with other frontier AI vendors, from an exposure management perspective. 

Specifically, Tenable seeks to determine and benchmark how frontier AI models’ advanced reasoning capabilities can enhance exposure management, including exposure analysis, asset intelligence, attack path mapping, vulnerability prioritization, risk-pattern recognition, and synchronized remediation. 

Tenable is using these insights to boost how it protects its own infrastructure, and to improve how we help our partners and customers protect their environments against AI-fueled cyber attacks.

Claude Mythos Preview’s closely-controlled, super-powerful cybersecurity capabilities will most likely one day become mainstream and widely available, including to bad actors who will undoubtedly try to misuse them for malicious purposes. When that happens, organizations need to be ready to withstand a vulnerability discovery tsunami, not to mention AI-driven attacks leveraging sophisticated agentic capabilities.

Of course, cyber defenders already face the challenge of defending their environments against AI-boosted cyber attacks. Even though they may not be as advanced as Claude Mythos Preview, there are agentic AI tools publicly available today that discover vulnerabilities at machine-speed and that, if used maliciously, can automate and accelerate exploit preparation.

And there is always the possibility that attackers will manage to get access to highly-advanced frontier AI models that haven’t been publicly released. Attackers routinely find ways to breach LLM security, so once specific advanced AI capabilities are developed, the potential exists for attackers to get their hands on them even before they become widely available and mainstream.

This is the new world that cybersecurity teams face: Attackers have access to increasingly powerful agentic AI tools for discovering exposures and deploying and operationalizing sophisticated cyber attacks at machine speed.

“Mythos Preview continues a long-term trend that we’ve been warning about for some time: within 6 to 12 months, we expect that many other AI companies will have Mythos-class models, and they could release them without safeguards that prevent misuse,” Anthropic wrote in a blog post in June 2026. “In that world, cyberattacks could occur much more often, and in much more unpredictable forms. It’s imperative that cyberdefenders adapt to maintain pace.”

The takeaway for cyber defenders: The machine-speed discovery and weaponization of zero-day vulnerabilities and of other security issues is a reality. The genie can’t be put back into the bottle.

Now more than ever, preventive cybersecurity is critical, so that you can remediate your most critical exposures at machine-speed before attackers armed with autonomous AI agents get a chance to exploit them. To accomplish this, AI-driven exposure management is key.

An AI-driven exposure management platform like Tenable One helps you to preemptively, continuously, and comprehensively detect unpatched vulnerabilities, misconfigurations, and overprivileged identities; surface the ones that represent the highest risk to your organization specifically; and coordinate their remediation autonomously, at machine-speed across your entire attack surface before attackers strike. 

5 steps to prepare for Mythos’ impact on cybersecurity

With an AI-driven exposure management platform, cyber defenders can take these five foundational steps for preemptive cybersecurity for the Claude Mythos Preview cybersecurity era.

• Ensure you have full asset visibility. It’s key to maintain a continuously updated inventory of all of your hybrid environment’s assets — approved and unapproved — including IT systems, operational technology (OT), AI tools, and cloud workloads. You also need to account for all of your assets’ security issues, such as unpatched vulnerabilities, misconfigurations, exposed secrets, and excessive identity permissions.
• Prioritize vulnerabilities based on precise business and technical context. Now that AI is exponentially increasing the number of vulnerability discoveries, you need to ruthlessly pinpoint the 1.6% of vulnerabilities that are most likely to put your organization at imminent risk. Otherwise, you’ll get overwhelmed by an unmanageable number of CVEs and other security issues, and fail to identify the exposures that put your organization at the highest risk.
• Use attack path analysis to spot toxic combinations of exposures. You must preemptively identify and defuse toxic combinations of preventable cybersecurity risks that malicious actors link together to launch cyber attacks, such as a low-severity CVE, a misconfigured cloud database, and an overprivileged identity, before an attacker unleashes an AI-boosted exploit to breach your network and move laterally.
• Adopt adversarial exposure validation (AEV). With AEV — a continuous loop of automated red teaming — you can assess your environment against the MITRE ATT&CK framework to learn how well-prepared you are to face AI exploits that leverage AI capabilities from frontier AI models.
• Use AI security to secure your AI systems. AI models, training pipelines, and highly-privileged autonomous agents have become highly attractive targets for cyber attackers. To protect these AI systems, you need an exposure management platform equipped with agentic AI engines, such as Tenable Hexa AI, to automate the preemptive detection, tagging, and remediation of exposures at Claude Mythos-like speed and scale.

Let's look at some of the ways in which exposure management is intrinsically desined to both help you prepare for and mitigate risks seen from frontier LLMs.

Preemptive, AI-driven exposure management

Implement comprehensive exposure management via platforms like Tenable One to gain unified visibility across IT, cloud, identities, and AI assets. This is designed to ensure hidden relationships, toxic exposure combinations, and risk patterns are mapped holistically before an attacker aided by AI can find them.

Strict identity and access governance

Treat frontier AI models like privileged corporate assets. Implement strict validation controls, enforce the principle of least privilege for internal AI agents, and establish explicit model permissions to prevent unauthorized data access or unapproved tool execution.

AI threat intelligence sharing

Actively monitor AI cybersecurity news regarding initiatives like Project Glasswing, OpenAI TAC (Trusted Access for Cyber) and OpenAI Daybreak, and participate in AI security consortiums. Leveraging early Project Glasswing software vulnerabilities research and participating in the drafting and refinement of AI security frameworks helps CISOs and their organizations benchmark their defenses against the most advanced exploitation methods emerging in the wild.

Continuous attack path analysis

Shift away from isolated CVE tracking and focus instead on continuous attack path analysis. By looking at infrastructure through an adversarial lens, defenders can identify critical choke points: the places where potential attack paths merge before leading to a critical asset. By prioritizing remediation of the most critical choke points, security teams can break complex, AI-generated exploit chains with precise, strategic fixes that block paths to the largest number of critical assets from threat actors.

Agentic AI defensive playbooks

Fight fire with machine-speed fire, leveraging AI for security. Leverage agentic AI remediation engines, such as Tenable Hexa AI custom agents built via the Model Context Protocol (MCP), to query your exposure data fabric in plain language. This allows defensive agents to dynamically discover threats, configure scans, and automate workflows at the exact pace the threat develops.

In conclusion, true preemptive cybersecurity in the age of AI requires establishing an exposure management program where exposure analysis, prioritization, and automated vulnerability remediation operate in a continuous, automated loop driven by agentic AI.

When deploying and operating a powerful frontier LLM like Claude Mythos, a CISO faces a unique challenge. Unlike typical enterprise software, an autonomous frontier model continuously ingests massive amounts of data, interacts with external codebases, and spins up its own internal workflows.

Traditional IT compliance frameworks can't handle this level of automation. To ensure continuous compliance with strict global data privacy regulations like the EU’s General Data Protection Regulation (GDPR) or the California Consumer Privacy Act (CCPA) without weakening innovation, CISOs must adapt modern, AI-specific security frameworks and risk management methodologies.

Core frameworks for AI data privacy and governance

NIST AI Risk Management Framework (AI RMF 1.0)

The standard NIST Cybersecurity Framework isn't enough. CISOs should implement the NIST AI Risk Management Framework (AI RMF 1.0), which is specifically designed to manage risks around AI data privacy, bias, and model trustworthiness.Its operational focus forces your team to map out how data flows into the model's training pipeline and context window, ensuring neither corporate intellectual property nor customers’ personally identifiable information (PII) is permanently absorbed by the model or leaked in telemetry.

ISO/IEC 42001 (Artificial Intelligence Management System)

As a leading international standard for AI governance, ISO 42001 provides an excellent baseline for corporate compliance.By establishing clear oversight parameters around system accountability, data handling transparency, and continuous risk assessments, ISO 42001 gives your board of directors auditable evidence that your LLM security controls are designed to meet global standards.

The MITRE ATLAS framework (Adversarial Threat Landscape for Artificial-Intelligence Systems)

MITRE ATLAS focuses on technical threats, helping your cyber teams map out specific AI threats and attack vectors — such as prompt injection, model poisoning, or data exfiltration — and zero-in on actively defending the specific data pipelines regulators care about most.

The Open Exposure Management Framework (OEMF)

The Open Exposure Management Framework (OEMF) seeks to address exposure management gaps in existing cybersecurity frameworks. Its benefits include providing cybersecurity professionals with a structured methodology to effectively prevent exploitable technology configurations at scale; more efficiently discovering, prioritizing and resolving exposures; and maximizing limited resources.