Why Cisco Is Changing Vulnerability Disclosure in the Age of AI

For decades, Cisco vulnerability management followed a familiar model.

A security vulnerability was discovered. A CVE was assigned. Cisco published an advisory describing the affected products, attack conditions, severity, workarounds, and fixed software releases. Enterprise customers then evaluated that information and decided how urgently they needed to upgrade.

That process gave network and security teams the technical information needed to answer important questions:

Does this vulnerability affect our environment?
Is the vulnerable feature enabled?
Can the vulnerability be exploited remotely?
Does exploitation require authentication or administrative access?
Can exposure be reduced through configuration or segmentation?
Does the security risk outweigh the operational risk of changing software?

Those are reasonable questions, especially for a large enterprise operating thousands of Cisco routers, switches, firewalls, wireless controllers, identity systems, collaboration platforms, and management appliances.

Cisco now believes the traditional vulnerability disclosure process cannot continue operating in exactly the same way as artificial intelligence dramatically increases the speed and scale of vulnerability discovery.

Cisco calls its new approach the Risk Based Vulnerability Disclosure Model.

This is more than a change to the Cisco Product Security Incident Response Team publication calendar. It represents a major change in how Cisco intends to discover, group, disclose, prioritize, and remediate vulnerabilities across its product portfolio.

To understand why Cisco is taking this path, we need to understand Claude Mythos Preview, Project Glasswing, and the cybersecurity problem that Cisco believes these technologies have exposed.

The Event That Changed Cisco’s Security Assumptions

In April 2026, Anthropic announced Claude Mythos Preview and Project Glasswing.

Claude Mythos Preview is a frontier artificial intelligence model designed with advanced cybersecurity capabilities. Anthropic states that the model can analyze software, identify previously unknown vulnerabilities, and, in many cases, develop working methods of exploiting them.

This is not limited to finding obvious coding mistakes or matching source code against a database of known insecure patterns.

According to Anthropic, Claude Mythos Preview found subtle vulnerabilities that had survived years of human code review, traditional automated testing, fuzzing, and security research.

Examples disclosed by Anthropic included:

A vulnerability in OpenBSD that had existed for approximately 27 years
A vulnerability in FFmpeg that had existed for approximately 16 years
Multiple Linux kernel vulnerabilities that could be combined into an exploitation chain
Vulnerabilities affecting major operating systems and web browsers

Anthropic reported that Claude Mythos Preview was able to identify many vulnerabilities and develop related exploits with little or no human guidance.

On Anthropic’s CyberGym vulnerability reproduction benchmark, Claude Mythos Preview achieved a reported score of 83.1 percent, compared with 66.6 percent for Claude Opus 4.6.

The significance is not simply that one model received a higher benchmark score.

The significant development is that frontier artificial intelligence is becoming competitive with highly skilled human security researchers at tasks that previously required extensive expertise, time, manual analysis, and experimentation.

Vulnerability research is beginning to move from human speed to machine speed.

What Is Project Glasswing?

Project Glasswing is an Anthropic initiative intended to secure critical software before advanced artificial intelligence capabilities become widely available to attackers.

The initiative provides selected infrastructure providers, software companies, security vendors, and open source organizations with early access to Claude Mythos Preview.

The purpose is to give defenders an opportunity to use the technology first.

Project Glasswing launch partners include:

Amazon Web Services
Anthropic
Apple
Broadcom
Cisco
CrowdStrike
Google
JPMorganChase
The Linux Foundation
Microsoft
NVIDIA
Palo Alto Networks

Anthropic also provided access to additional organizations responsible for maintaining important software and open source infrastructure.

By May 2026, Anthropic stated that approximately 50 participating organizations had used Claude Mythos Preview to find more than 10,000 vulnerabilities rated as high or critical across systemically important software.

Project Glasswing is therefore much larger than a research experiment involving one artificial intelligence company.

It is an organized attempt to identify and remediate vulnerabilities throughout critical global infrastructure before comparable offensive capabilities become inexpensive, accessible, and widely distributed.

Why Cisco Joined Project Glasswing

Cisco operates one of the largest and most consequential software and hardware portfolios in enterprise technology.

Cisco products carry traffic through corporate networks, data centers, service provider backbones, cloud environments, hospitals, universities, financial institutions, government agencies, industrial facilities, and critical infrastructure.

This makes Cisco both an important defender and an extremely valuable target.

Cisco joined Project Glasswing because it believes frontier artificial intelligence has crossed a threshold that fundamentally changes the security problem.

Cisco’s Chief Security and Trust Officer stated that the company’s work with these models demonstrated that vulnerabilities across hardware and software could be identified and fixed at a pace and scale that had previously been impossible.

The company’s conclusion is that traditional methods of hardening software are no longer sufficient by themselves.

Human security researchers, static analysis, fuzzing, penetration testing, code review, and bug bounty programs will remain important. However, Cisco now has evidence that advanced artificial intelligence can identify entire categories of vulnerabilities that those processes did not discover.

Cisco is therefore using Project Glasswing and Claude Mythos Preview to examine foundational software, identify weaknesses, validate findings, and strengthen products before the same level of capability becomes commonly available to attackers.

How Cisco Is Using Claude Mythos

It is important not to oversimplify Cisco’s use of Claude Mythos.

Cisco is not merely opening a chatbot and asking it whether IOS XE contains vulnerabilities.

Frontier cybersecurity models can be incorporated into structured security research processes that provide the model with code, system behavior, configuration context, testing environments, and clearly defined objectives.

Within such a process, an advanced model may assist researchers with tasks including:

Reviewing large software codebases
Identifying suspicious logic and unsafe assumptions
Analyzing memory handling and input validation
Examining authentication and authorization flows
Reviewing configuration dependent behavior
Testing live systems in controlled environments
Reproducing suspected vulnerabilities
Developing proof of concept exploitation methods
Determining whether multiple weaknesses can be combined
Searching for similar weaknesses in related products
Helping engineers understand root cause
Assisting with remediation and verification

The model can search much more broadly than a human researcher working on one individual defect.

For example, after identifying an unsafe input validation pattern in one component, an artificial intelligence system may search other components and products for similar implementations.

This can transform one vulnerability investigation into the discovery of an entire vulnerability class.

That is one reason Cisco is moving toward software hardening releases and grouped vulnerability remediation instead of treating every defect as a completely isolated event.

The model does not eliminate the need for Cisco security engineers.

Human experts still need to determine whether a finding is legitimate, whether it is exploitable in a supported configuration, what products are affected, what the security impact is, how the code should be corrected, whether the correction creates regressions, and how the issue should be disclosed.

The major change is scale.

Artificial intelligence allows Cisco’s researchers to investigate more code, test more conditions, correlate more findings, and identify more potential vulnerabilities than a human team could reasonably process through traditional manual methods alone.

Project Glasswing Is a Defensive Head Start

Project Glasswing is based on a simple and uncomfortable assumption.

The capabilities available to defenders today will eventually become available to attackers.

Anthropic is giving Cisco and the other Project Glasswing partners early access to Claude Mythos Preview so they can search for vulnerabilities before advanced cyber capabilities become more broadly available.

This does not create a permanent advantage.

It creates a temporary opportunity.

Cisco can use that opportunity to identify vulnerabilities, develop corrections, harden software releases, and encourage customers to deploy those releases before attackers gain equivalent capability.

That is why Cisco describes the current situation with a sense of urgency.

The company does not believe it has many years to prepare. Cisco believes the necessary changes must happen within months because frontier artificial intelligence capabilities are improving extremely quickly.

The Problem Is Not Limited to Claude

One question immediately comes to mind.

If Claude Mythos can discover these vulnerabilities, what makes anyone believe that Claude will remain the only model capable of doing it?

The answer is that it will not.

The United States has several organizations developing frontier artificial intelligence systems, including:

Anthropic
OpenAI
Google
Meta
xAI

China has also produced highly capable models and artificial intelligence research organizations, including:

DeepSeek
Alibaba Qwen
Moonshot AI and Kimi
Tencent Hunyuan
Baidu ERNIE
Zhipu AI

Chinese artificial intelligence models should not be dismissed as inferior copies of American systems.

They have become highly competitive in coding, reasoning, mathematics, model efficiency, agentic workflows, and open model development.

Some researchers, analysts, and technology leaders believe China has matched or surpassed the United States in particular areas of artificial intelligence development.

Whether China has surpassed the United States overall remains debatable because artificial intelligence leadership depends on what is being measured. Different organizations lead in model performance, computing infrastructure, chip development, research output, deployment scale, investment, efficiency, and open source adoption.

From a cybersecurity perspective, deciding which country holds the overall lead is not the most important question.

The more important fact is this:

World class artificial intelligence capability is no longer limited to one company, one model, or one country.

If Anthropic can build a model that accelerates vulnerability discovery, it is reasonable to expect OpenAI, Google, Meta, xAI, DeepSeek, Qwen, Kimi, and future models to develop similar capabilities.

Those capabilities may also be developed by governments, intelligence agencies, military organizations, universities, cybercriminal groups, and private research laboratories.

The better strategic question is no longer:

Can Claude discover vulnerabilities?

The better question is:

How long will it take before multiple advanced artificial intelligence systems independently discover the same vulnerability?

How This Connects to Cisco’s Disclosure Strategy

Cisco’s use of Claude Mythos creates a vulnerability management problem as well as an opportunity.

If artificial intelligence allows Cisco to discover vulnerabilities at unprecedented scale, the company must determine how to process, correct, document, prioritize, and disclose that volume of findings.

The traditional process was designed around a much smaller number of vulnerabilities.

A vulnerability would be assigned its own CVE, investigated individually, documented in a standalone advisory, scored, published, and tracked through remediation.

That model becomes difficult to sustain if advanced artificial intelligence begins identifying thousands of related weaknesses across multiple products.

The CVE system was not designed for a world where autonomous models can continuously examine enormous codebases and discover entire classes of defects.

Cisco’s response is its Risk Based Vulnerability Disclosure Model.

The strategy includes:

A predictable disclosure calendar
Twice monthly publication windows
Advance customer notifications
Security hardening releases
Bundled or umbrella CVEs
Greater prioritization of high risk findings
Less granular disclosure for some lower risk findings

The new model attempts to move customers away from reacting to every defect as a separate emergency and toward regularly deploying hardened and supported software.

What Cisco Is Changing

Scheduled Security Disclosures

Cisco plans to publish scheduled security advisories and hardening releases on the first and third Wednesday of each month when the corresponding software is available.

Critical vulnerabilities, active exploitation, serious third party issues, and other urgent conditions can still result in publications outside the normal schedule.

The purpose of the scheduled cadence is to make vulnerability management more predictable.

Instead of responding to an unpredictable stream of individual disclosures, enterprises can align Cisco security releases with lab testing, change control, maintenance planning, and deployment processes.

Seven Day Advance Notifications

Cisco plans to provide advance notification approximately seven days before a scheduled disclosure date.

The notification identifies affected technologies or product families. It may not reveal the complete vulnerability details, but it gives customers time to reserve engineering resources, prepare test environments, begin change control, and identify potential maintenance windows.

For a large enterprise, seven days may not be enough time to test and deploy a new network operating system release. However, advance notice is still more useful than learning about the release only after publication.

Security Hardening Releases

Cisco is introducing security hardening releases that correct multiple internally discovered vulnerabilities within a product family.

Instead of every defect producing a separate patching event, Cisco can remediate groups of vulnerabilities through a hardened software release.

Cisco expects major network operating system platforms to receive security updates on a quarterly basis, although releases may occur outside that cadence when conditions require it.

Umbrella CVEs

Cisco is also changing how some vulnerabilities are represented through the CVE system.

Vulnerabilities that share a common weakness category may be grouped under an umbrella CVE.

For example, several defects involving the same class of input validation weakness could be grouped under one CVE associated with the relevant Common Weakness Enumeration category.

The assigned severity would reflect the highest severity represented within the grouped vulnerabilities.

This means customers may not receive an individual CVE and a complete technical explanation for every internally discovered defect included in a hardening release.

Less Detail for Certain Lower Risk Findings

This is the part of the strategy that enterprise Cisco customers need to understand most clearly.

Some lower risk vulnerabilities, including some internally discovered vulnerabilities, may receive less detailed public disclosure.

Lower risk findings may be addressed through hardening releases and summary information rather than through individual advisories containing detailed technical descriptions.

For vulnerabilities that are critical, actively exploited, highly likely to be exploited, or require direct defensive action, Cisco states that it will continue providing detailed information and individual advisories where appropriate.

Is Cisco Hiding Vulnerabilities?

The answer depends on what someone means by hiding.

Cisco is not saying that it will secretly modify software and provide customers with no information.

Customers will still receive fixed software, affected product information, release guidance, security notifications, CVE information where applicable, and summary details about hardening releases.

However, Cisco may intentionally avoid publishing the exact technical root cause, affected code path, exploitation method, and detailed conditions for every lower risk vulnerability.

From an enterprise customer’s perspective, that is a reduction in technical transparency.

The practical message may become:

This software release contains important security hardening. Cisco recommends that you test and deploy it, even though every underlying vulnerability will not receive its own detailed public advisory.

That is a major change from a vulnerability management process based entirely on evaluating individual CVEs.

Why Detailed Advisories Can Help Attackers

A detailed Cisco advisory can provide attackers with valuable starting information.

It may reveal:

The vulnerable product
The affected feature or protocol
The required access level
The triggering input or condition
The affected software versions
The release containing the correction
Available workarounds

An attacker can then obtain vulnerable and fixed software, compare the versions, identify the corrected component, reproduce the condition, and develop an exploit.

This process existed long before generative artificial intelligence.

Artificial intelligence can accelerate portions of it.

A capable model may help analyze binaries, compare code paths, correlate release information, identify changed logic, generate test inputs, reproduce failures, and refine exploitation techniques.

Meanwhile, enterprise customers may require weeks or months to qualify a new IOS XE, NX OS, IOS XR, ASA, Secure Firewall, ISE, or collaboration platform release.

They must validate routing, switching, redundancy, authentication, firewall policy, application dependencies, hardware compatibility, licensing, monitoring, automation, and management integrations.

This creates a dangerous timing imbalance.

Attackers may be able to analyze and weaponize a disclosed vulnerability faster than a large enterprise can test and deploy the fixed software.

Cisco believes the time between vulnerability disclosure and practical exploitation is collapsing.

Its new strategy attempts to reduce the amount of immediately useful technical information available to attackers while making corrected software available to customers.

What Limited Disclosure Can Accomplish

Reducing public technical detail cannot permanently keep a vulnerability secret.

Other researchers may independently discover it. Attackers may have their own models. Nation state organizations may already know about a weakness. Researchers can compare old and new software even when no detailed advisory is available.

Limited disclosure is therefore not a complete security control.

It is a strategy for buying time.

By withholding some granular details, Cisco may avoid giving every attacker an immediate and organized roadmap.

This will not stop the most capable adversaries, but it may delay opportunistic attackers, automated exploit developers, and organizations that depend heavily on public advisories to identify targets.

Even a limited delay can be valuable when enterprises need time to test and deploy software across globally distributed infrastructure.

What This Means for Cisco Customers

CVE Based Decisions Will Become More Difficult

Many network and security teams currently evaluate each Cisco advisory independently.

For example, an organization may determine that a vulnerability affecting a web management interface does not create meaningful exposure because the service is disabled or restricted to a protected management network.

That type of technical exposure analysis remains valid when sufficient information is available.

However, when multiple internally discovered vulnerabilities are grouped into a hardening release with limited details, customers may not have enough information to prove that every underlying issue is not applicable.

This does not eliminate the enterprise’s authority to make risk based decisions.

It means those decisions may need to be made with less vulnerability specific evidence.

Software Currency Becomes Part of the Security Baseline

Cisco wants customers running supported and security hardened software instead of remaining on older releases through a collection of individual workarounds.

An older release may have no publicly documented critical vulnerability and still carry greater security risk because it predates multiple bundled or undisclosed hardening corrections.

Under this model, the absence of a publicly documented CVE does not necessarily mean that an older release has no known security weaknesses.

Testing Must Become Faster and More Repeatable

Enterprises should not blindly install every Cisco software release directly into production.

Network software upgrades can introduce serious operational consequences. A security correction does not remove the need to test routing protocols, high availability, control plane behavior, hardware compatibility, authentication, licensing, automation, monitoring, and rollback procedures.

The correct response is not blind patching.

The correct response is developing a mature and repeatable process that can evaluate and deploy Cisco software more efficiently.

That process should include:

Representative laboratory environments
Automated validation before and after upgrades
Approved software baselines for each platform
Review of Cisco release notes and known defects
Documented rollback procedures
High availability and failover testing
Monitoring of Cisco advance notifications
Coordination among network engineering, operations, and information security
Formal risk acceptance when an upgrade must be deferred

This Does Not Eliminate Risk Based Patch Management

One possible internal interpretation of Cisco’s new strategy is:

Cisco says the release is hardened, so install it without further analysis.

That would be an overcorrection.

Infrastructure teams must still evaluate:

The stability history of the target release
Known defects and field notices
Hardware and feature compatibility
The operational impact of the upgrade
Rollback capabilities
The criticality of the affected environment
The exposure of the affected platform

Information security should own or formally govern the security risk decision.

Network engineering should provide the technical evidence concerning configuration, exposure, feature usage, software behavior, operational dependencies, and upgrade risk.

Neither group should make the decision in isolation.

Cisco’s new approach changes the information available to that process, but it does not remove the need for accountable risk management.

My View of Cisco’s Direction

I understand why Cisco is taking this path.

Publishing a detailed technical explanation for every internally discovered vulnerability may no longer provide the same defensive value it once did.

In an environment where artificial intelligence can accelerate patch analysis and exploit development, a detailed advisory can also become a roadmap for attackers.

Cisco is attempting to replace an unpredictable vulnerability by vulnerability process with a systematic hardening model built around scheduled releases, grouped remediation, and prioritized disclosure.

The strategy makes sense in principle.

However, its success will depend on Cisco’s execution.

Cisco must provide customers with:

Stable and thoroughly tested hardened releases
Clear product and release mappings
Reliable advance notifications
Accurate upgrade guidance
Useful summary information
Transparent handling of critical and actively exploited vulnerabilities
Practical tools for understanding the security state of installed products

If Cisco provides less vulnerability detail but does not significantly improve software quality, release stability, and upgrade predictability, large enterprises will struggle to adopt the new model.

Trust must work in both directions.

Cisco is asking customers to place greater trust in release level security assurance.

In return, customers must be able to trust that a security hardening release is operationally mature enough to deploy across critical infrastructure.

Final Thoughts

Cisco’s Risk Based Vulnerability Disclosure Model is not based on the belief that Claude Mythos is the only artificial intelligence system capable of discovering software vulnerabilities.

I believe Cisco is preparing for exactly the opposite reality.

Project Glasswing gives Cisco a temporary defensive advantage by providing early access to Claude Mythos Preview.

Cisco can use that advantage to search its products, identify vulnerability classes, correct software, and deliver hardened releases before comparable cyber capabilities become commonly available.

However, Anthropic does not have a permanent monopoly on advanced artificial intelligence.

OpenAI, Google, Meta, xAI, DeepSeek, Qwen, Kimi, and future systems are improving rapidly.

Some of the strongest Chinese models already compete directly with leading American systems. Similar capabilities may also emerge from private laboratories, governments, universities, intelligence organizations, and open source communities.

Artificial intelligence assisted vulnerability discovery is becoming a global capability.

Cisco cannot prevent every independent discovery, and it cannot permanently conceal the differences between vulnerable and corrected software.

What Cisco can do is avoid immediately providing every attacker with a detailed technical roadmap, release hardened software on a predictable schedule, and give customers an opportunity to deploy corrections before vulnerabilities are widely weaponized.

The industry is moving from a model centered almost entirely on individual CVEs toward one increasingly centered on software currency, systematic hardening, and deployment speed.

For Cisco customers, the lesson is not that we should patch blindly.

The lesson is that we need stronger processes for evaluating, testing, and deploying network software because the amount of time available to make those decisions is shrinking.

The future of vulnerability management will not be determined solely by which company or country has the most capable artificial intelligence.

It will be determined by whether defenders can discover, validate, remediate, test, and deploy corrections faster than attackers can turn software weaknesses into working exploits.

Cisco’s New Vulnerability Disclosure Strategy in the Age of AI