Cybersecurity has historically been constrained by one scarce resource: expert attention. Modern software contains millions of lines of code, enormous dependency trees and constantly changing infrastructure, while experienced vulnerability researchers can deeply investigate only a fraction of that attack surface.

Frontier AI is beginning to change that equation.

Anthropic’s Claude Mythos line represents one of the clearest examples of how advanced AI models can be applied to vulnerability discovery and defensive cybersecurity. Anthropic introduced Claude Mythos Preview and Project Glasswing on April 7, 2026, describing Glasswing as an initiative intended to put advanced cyber capabilities to work securing critical software.

The significance of this development extends beyond one model or one company. If AI can inspect code, generate vulnerability hypotheses, investigate candidate weaknesses and assist with remediation at substantially greater scale than conventional security teams, then some of the basic economics of vulnerability research begin to change.

Finding bugs becomes faster.

But that creates a new bottleneck: determining which findings are real, coordinating disclosure, building patches and deploying remediation before vulnerabilities become useful to attackers. Anthropic’s early Glasswing results explicitly describe this shift from discovery toward verification, disclosure and patching as the limiting part of the process.

That may ultimately be the most important lesson from Mythos and Project Glasswing for defensive cybersecurity.

What Is Claude Mythos?

Claude Mythos is Anthropic’s high-capability model line designed for trusted access to unusually powerful cybersecurity capabilities. The first publicly described version, Claude Mythos Preview, was introduced in April 2026 as an unreleased frontier model with particularly strong computer-security capabilities. Anthropic said its internal testing showed that Mythos Preview had reached a level where it could outperform all but highly skilled human specialists at some vulnerability discovery and exploitation tasks.

That description needs an important qualification.

Mythos should not be understood simply as an automated vulnerability scanner.

Traditional scanners usually search for known signatures, unsafe patterns or predetermined vulnerability classes. Frontier models can potentially participate in a much broader reasoning process: understanding unfamiliar code, following relationships between components, identifying suspicious assumptions and investigating whether those assumptions can become exploitable.

That is much closer to the workflow of a security researcher.

It is also why Anthropic has treated access more cautiously than it would access to an ordinary general-purpose model. As of August 19, 2026, Anthropic describes Mythos 5 as available through a limited trusted-access model rather than ordinary unrestricted public access. Anthropic’s current Mythos page says access was restored on July 1 to a set of U.S. organizations following an earlier temporary suspension, while the company continues operating access through trusted cybersecurity programs.

The distinction matters because advanced cyber capability is inherently dual use.

The same model that helps defenders identify a previously unknown vulnerability can potentially help an attacker understand how that vulnerability might be exploited.

What Is Project Glasswing?

Project Glasswing is Anthropic’s defensive cybersecurity initiative built around using powerful AI capabilities to help secure important software before similar capabilities become widely available to malicious actors.

Anthropic launched the project in April 2026 with an initial group that included major technology, infrastructure and security organizations. The project was explicitly framed around using Claude Mythos Preview to identify and remediate software vulnerabilities.

The project is important because it provides a real-world environment for answering a question that benchmarks alone cannot answer:

What happens when frontier cybersecurity models are applied to large production codebases at scale?

Synthetic security benchmarks can measure specific technical capabilities.

Actual software introduces a much harder environment.

Codebases are enormous.

Documentation is incomplete.

Legacy assumptions accumulate over years.

Dependencies interact unexpectedly.

A theoretically suspicious function may never be reachable by an attacker.

Another piece of code may look harmless until combined with several other components.

Project Glasswing moves AI cybersecurity from isolated evaluation toward real vulnerability research.

From Mythos Preview to Mythos 5

The Mythos program progressed rapidly after the April launch.

On June 9, 2026, Anthropic announced Claude Mythos 5, an updated model intended initially for a small group of cyberdefenders and infrastructure providers through Project Glasswing and subsequent trusted-access programs. Anthropic described Mythos 5 as using the same underlying model as Fable 5 but with some cybersecurity safeguards removed for approved defensive users.

This architecture illustrates an important question for future cybersecurity AI.

A model can possess highly valuable defensive capabilities while those same capabilities create substantial misuse potential.

One response is to make the underlying capability broadly available and attempt to classify individual requests.

Another is to maintain stronger safeguards for general users while providing more capable configurations only to vetted security organizations.

Anthropic has largely pursued the latter model for Mythos-level cybersecurity access. Its current Mythos information describes a small trusted-access group rather than ordinary open availability.

That approach is likely to remain an important part of the broader discussion around frontier cyber models.

Why Frontier AI Matters for Vulnerability Discovery

Vulnerability discovery is difficult because software bugs are rarely labeled as vulnerabilities.

A researcher needs to determine how the program is intended to work, how it actually works and whether an attacker can create a meaningful difference between the two.

That often requires reasoning across multiple functions or components.

A parser trusts a value because another component was supposed to validate it.

An authorization check exists but does not cover every execution path.

A memory-management assumption holds under ordinary input but fails under an unusual sequence of operations.

A privilege boundary appears secure when one function is inspected but can be bypassed through a different workflow.

These are not simply pattern-matching problems.

They require contextual reasoning.

Frontier models become interesting for cybersecurity when they can participate in that deeper analysis.

AI Can Explore More Code Than Human Researchers Alone

A security researcher may be exceptionally skilled but still has finite time.

Large organizations maintain huge volumes of software.

Open-source ecosystems contain millions of projects.

Security-sensitive dependencies may remain largely unreviewed for years simply because nobody has enough time to inspect them deeply.

AI changes the potential scale of that process.

A model can inspect many candidate code paths, generate possible vulnerability hypotheses and present promising cases for deeper validation.

Project Glasswing’s early results illustrate the scale involved. Anthropic reported in May that it and roughly 50 initial partners had used Mythos Preview to surface more than 10,000 candidate high- or critical-severity vulnerabilities across important software. Anthropic simultaneously emphasized that human verification and remediation had become the limiting factor.

The significance of those numbers is not that every AI-produced result should be trusted.

It is almost the opposite.

The more quickly AI can generate plausible findings, the more important rigorous human triage becomes.

Candidate Vulnerabilities Are Not Confirmed Vulnerabilities

This distinction is central to understanding responsible AI-assisted security research.

A language model can identify something suspicious.

That does not automatically mean a real attacker can exploit it.

Perhaps the affected function is unreachable from an external interface.

Perhaps another authorization layer blocks the attack.

Perhaps configuration makes the vulnerable condition impossible.

Perhaps the model simply misunderstood the implementation.

Anthropic’s coordinated vulnerability disclosure process explicitly uses human triage and review before candidate Mythos findings are reported to maintainers. Its public disclosure dashboard states that candidate vulnerabilities are reviewed by external security researchers and that only human-reviewed high- or critical-severity issues move through coordinated disclosure.

This human validation layer is one of the most important patterns for responsible AI vulnerability research.

AI expands discovery.

Security researchers establish reality.

AI-Assisted Vulnerability Discovery Changes the Bottleneck

For decades, cybersecurity has often struggled to find enough weaknesses before attackers do.

Frontier models may gradually change that problem.

If a capable AI can surface large numbers of credible vulnerability candidates quickly, the limiting resource becomes the organization’s ability to investigate them.

Anthropic’s Project Glasswing update explicitly identifies verifying, disclosing and patching vulnerabilities as the emerging bottleneck rather than initial discovery.

That has important implications for security operations.

A company might dramatically improve vulnerability discovery while gaining little practical security if remediation remains slow.

An open-source maintainer may receive hundreds of technically sophisticated reports but lack engineering resources to investigate them.

A security team may suddenly have much greater visibility into software weaknesses while lacking a process for deciding what should be fixed first.

AI can therefore create an unusual problem:

better discovery can initially increase operational security pressure.

Verification Becomes More Valuable as Discovery Gets Cheaper

If AI can generate candidate vulnerabilities at increasing scale, skilled validation becomes a higher-value activity.

Security teams need researchers capable of determining whether findings are reachable, reproducible and exploitable.

They need to understand prerequisites.

They need to distinguish theoretical weaknesses from realistic attack paths.

They need to evaluate actual business impact.

They also need to reject false positives efficiently.

This resembles what is already happening with traditional security tooling, but potentially at a much larger scale.

Security programs that simply count AI-generated vulnerabilities may become overwhelmed.

Programs that build strong triage pipelines can convert increased discovery capacity into actual risk reduction.

Coordinated Vulnerability Disclosure Becomes More Important

Faster vulnerability discovery creates another challenge: disclosure.

A previously unknown vulnerability is valuable information.

Publishing it immediately can expose users before a patch exists.

Keeping it private indefinitely may leave users vulnerable without giving maintainers an opportunity to fix the problem.

Coordinated vulnerability disclosure exists to manage this tension.

Anthropic has created a specific disclosure framework for vulnerabilities discovered through its AI research and maintains a public dashboard tracking findings that have progressed through the disclosure process.

This is likely to become increasingly important as AI-generated vulnerability volume increases.

If a model can surface hundreds or thousands of candidate vulnerabilities, disclosure cannot remain an improvised email process.

Security infrastructure needs to scale alongside discovery.

Frontier AI Can Also Assist With Patching

Finding vulnerabilities is only half of the defensive problem.

Software needs to be fixed.

Anthropic reports that Project Glasswing partners have been using Mythos Preview not only to find vulnerabilities but also to help write patches and perform pre-release security checks. Anthropic has described vulnerability remediation as one of the major areas where it expects powerful cyber models to provide defensive value.

This is strategically important.

If AI accelerates only offensive analysis, defenders face increased pressure.

If it also accelerates remediation, some of that pressure can be offset.

The ideal workflow is therefore not:

AI discovers thousands of vulnerabilities.

The engineering team receives thousands of tickets.

Instead:

AI discovers a candidate.

A researcher validates the vulnerability.

AI assists with understanding the affected code.

Engineering develops a fix.

AI helps inspect related code paths.

The security team retests the attack path.

That creates a closed defensive loop.

AI Can Move Security Earlier in Development

Perhaps the most valuable use of frontier cybersecurity AI is not finding vulnerabilities in deployed software.

It is finding them before deployment.

Anthropic says Project Glasswing partners have used Mythos Preview for pre-release checks intended to identify vulnerabilities before they reach production.

That approach can fundamentally improve the economics of remediation.

Changing architecture before release is often much easier than fixing it after customers depend on the behavior.

A vulnerability discovered before deployment may require a code change.

The same vulnerability discovered after launch can require emergency patches, customer communication, incident investigation and production migration.

AI-assisted pre-release review can therefore complement traditional secure development rather than simply creating another post-release scanner.

AI-Assisted Penetration Testing

Frontier models also have clear applications inside authorized penetration testing.

Anthropic explicitly lists penetration testing among the defensive activities for which Mythos-class capabilities can be useful.

The model can help researchers understand an application, identify suspicious attack surfaces and explore potential relationships between findings.

But the Rules of Engagement do not change because AI is involved.

A penetration test still requires authorization.

Scope still matters.

Third-party systems remain outside scope unless explicitly approved.

Production safety remains important.

The tester still needs to determine when enough evidence has been collected.

AI changes the speed and breadth of analysis.

It does not expand the legal or operational permission to test.

This is the same principle behind ShabuShabu’s Security Testing Terms and testing methodology.

Why Mythos-Class Cybersecurity Is Dual Use

Cybersecurity is unusual because many of the skills required for defense closely resemble those required for attack.

Understanding a vulnerability helps defenders patch it.

The same understanding helps an attacker exploit it.

Knowing how to move through a compromised environment helps defenders reproduce attack paths.

It can also help offensive operators.

Anthropic has explicitly cited this dual-use problem as the reason Mythos-level capabilities are not simply exposed without restrictions. Its current materials describe Mythos 5 as a particularly powerful cyber model made available through trusted-access mechanisms because of the risk of malicious use.

This tension will likely affect the entire cybersecurity AI industry.

More Capable Models Need More Than Model-Level Safeguards

As cyber capability grows, safety cannot depend on a single refusal behavior.

Model safeguards matter.

Access control matters.

User verification matters.

Monitoring matters.

Capability restrictions matter.

Disclosure processes matter.

Anthropic’s June 30 update describes this explicitly as a defense-in-depth problem. The company says its general-purpose Fable 5 uses multiple cybersecurity safeguards, while Mythos 5 provides less-restricted capabilities only to selected trusted organizations.

This is analogous to privileged security tooling.

Organizations do not make production administrator credentials publicly available and rely solely on users behaving responsibly.

Capabilities themselves require governance.

The Temporary Mythos 5 Suspension Shows Why Deployment Policy Matters

The development of Mythos 5 also illustrates how quickly frontier-model deployment conditions can change.

On June 12, 2026, U.S. export controls caused Anthropic to suspend access to both Fable 5 and Mythos 5. After those controls were lifted, Anthropic restored Mythos 5 access on July 1 for an approved set of U.S. organizations while continuing to work toward broader access for Glasswing partners.

For organizations following frontier cybersecurity AI, this is a useful reminder that capability availability is not determined only by technical readiness.

Security policy, safeguards, trusted-access requirements and regulatory decisions can also affect deployment.

As of August 19, 2026, it would therefore be inaccurate to describe Mythos 5 as an unrestricted public cybersecurity model. Anthropic’s own current product page describes it as available to a limited set of trusted cybersecurity users.

What Project Glasswing Says About the Future of Security Research

The most interesting part of Glasswing may not be the specific organizations or models involved.

It is the workflow the project is beginning to reveal.

Large-scale AI systems identify candidate weaknesses.

Human researchers validate them.

Maintainers receive coordinated disclosures.

AI assists with patches.

Security teams perform pre-release checks.

The model becomes part of a broader security infrastructure rather than an isolated vulnerability scanner.

Anthropic has said its longer-term Glasswing goal is to support the security industry with models, tools, infrastructure and eventually stronger systems for vulnerability review and patching.

That is a much bigger ambition than simply building a better scanner.

The Future Security Problem May Be Vulnerability Throughput

Security organizations traditionally worry about detection coverage.

How many vulnerabilities are we missing?

Frontier AI introduces another question:

How many vulnerabilities can we realistically process?

Suppose AI can surface thousands of plausible findings.

Security researchers still need to triage them.

Maintainers need to understand them.

Developers need to patch them.

Users need to deploy those patches.

Adding discovery capacity without improving the rest of the security pipeline can simply move the bottleneck downstream.

Anthropic’s Glasswing updates repeatedly emphasize this point: AI has increased the rate at which vulnerabilities can be surfaced, placing greater pressure on human validation and patching systems.

The defensive-security industry therefore needs automation throughout the vulnerability lifecycle, not only at the discovery stage.

Human Researchers Become More Important, Not Less

At first glance, extremely capable vulnerability-discovery models may appear to reduce the need for security researchers.

In practice, they change the researcher’s role.

When candidate generation is scarce, researchers spend much of their time searching.

When candidate generation becomes abundant, researchers spend more time validating.

Is the vulnerability real?

Can an attacker reach it?

What permissions are required?

Does exploitation affect another user?

What is the minimum evidence necessary to prove impact?

How should the issue be disclosed?

Does the patch actually remove the attack path?

These decisions require security judgment.

Anthropic’s own disclosure workflow explicitly retains independent human triage for Mythos-discovered vulnerabilities rather than treating model output as automatically confirmed.

That is a strong indicator of where professional vulnerability research is heading.

AI Does Not Eliminate False Positives

Greater reasoning ability does not make a model infallible.

Frontier models can misunderstand code.

They can infer nonexistent attacker control.

They can miss configuration restrictions.

They can produce plausible explanations for impossible attack paths.

This is why raw model output should not automatically enter a customer-facing security report.

A professional penetration test still needs technical reproduction.

For ShabuShabu, that means the same validation standard should apply whether a candidate weakness was first identified manually, through conventional automation or with AI assistance.

The discovery mechanism changes.

The evidentiary standard should not.

Severity Still Depends on Real Impact

A model may identify an unusual security condition.

Severity depends on what the condition enables.

A bug reachable only from a highly privileged local administrator has a different risk profile from one accessible remotely without authentication.

A technically interesting information disclosure may reveal nothing sensitive.

A subtle authorization weakness may expose another customer’s entire account.

AI can assist with reasoning about those conditions.

But vulnerability severity ultimately belongs to the application context.

That is why ShabuShabu’s Security Crash Test model focuses on validated attack paths rather than raw vulnerability counts.

Frontier AI Can Strengthen Open-Source Security

Open-source software presents an especially important defensive opportunity.

A single widely used component may sit inside thousands or millions of downstream products while the maintainers responsible for it have comparatively limited security resources.

Project Glasswing was explicitly designed around important and widely deployed software, and Anthropic’s public disclosure program includes vulnerabilities discovered across open-source projects.

AI can help increase the amount of expert-level analysis available to these projects.

But again, discovery alone is not enough.

Maintainers need usable reports.

They need patches.

They need enough time for users to update.

The future of AI-assisted open-source security therefore depends as much on remediation infrastructure as on model capability.

Frontier AI May Shorten the Defender’s Time Window

Powerful AI does not remain a permanent advantage for one side.

Anthropic’s rationale for Glasswing is explicitly based on the expectation that Mythos-class cyber capabilities will become more common, making it important for defenders to adapt early.

If vulnerability analysis becomes cheaper and faster, attackers may be able to investigate newly released software or patches much more quickly.

This increases the value of secure development and pre-release testing.

Organizations should not assume complexity will protect vulnerabilities simply because exploitation once required highly specialized manual analysis.

When AI reduces the cost of technical reasoning, obscurity created by complexity becomes weaker.

AI Makes Secure Software Development More Important

One possible reaction to highly capable cybersecurity AI is to rely on AI to find vulnerabilities after developers create them.

That would be a mistake.

The better strategy is to combine AI-assisted detection with better engineering.

Authorization should be designed correctly.

Sensitive operations should use least privilege.

Input should be handled safely.

Security boundaries should be explicit.

Dependencies should be maintained.

High-risk releases should receive adversarial testing before deployment.

AI can increase the chance that mistakes are discovered.

Good architecture reduces the number of mistakes that need to be discovered.

The two approaches reinforce each other.

Project Glasswing Shows That Patching Must Scale

The vulnerability-discovery problem is already beginning to intersect with a software-maintenance problem.

Anthropic says it is working with third parties on ways to scale reviewing and patching of open-source vulnerabilities and improve how vulnerability information is communicated to maintainers.

That direction is important.

A future system may not simply report:

“Here is a vulnerability.”

It may provide validated analysis, identify related code, suggest a remediation, generate candidate patches, run tests and prepare enough evidence for a human maintainer to make a decision.

Security professionals still control the final process.

But the amount of mechanical work around each vulnerability can fall substantially.

AI Cybersecurity Will Increasingly Become Continuous

Traditional security assessments often occur at specific moments.

A penetration test before launch.

A security review before a major release.

A vulnerability scan on a schedule.

AI makes continuous reasoning more practical.

New code can be inspected when it is committed.

Potential security-sensitive changes can receive deeper analysis.

Known vulnerable patterns can be compared with nearby implementation logic.

Previously identified vulnerability families can be searched across the codebase.

The result is a security process where automated reasoning occurs continuously while deeper human validation is concentrated on the most important candidates.

Project Glasswing’s emphasis on both pre-release review and vulnerability remediation points toward this more continuous defensive model.

Mythos Does Not Make Traditional Penetration Testing Obsolete

A powerful code-analysis model still does not know everything about a production system.

Runtime configuration matters.

Authentication state matters.

Cloud architecture matters.

Real user permissions matter.

Business logic matters.

Third-party integrations matter.

Infrastructure differences between development and production matter.

A vulnerability that looks critical in source code may be unreachable in the actual environment.

A vulnerability invisible from source inspection may emerge through the interaction between several deployed systems.

This is why AI-assisted vulnerability discovery and traditional penetration testing services should be combined rather than treated as substitutes.

AI improves research coverage.

Runtime testing supplies operational truth.

AI-Assisted Defense Needs Responsible Disclosure

Powerful vulnerability discovery creates responsibility.

A researcher discovering a previously unknown weakness should not publish technical details before maintainers have a realistic opportunity to fix the issue.

Anthropic’s coordinated disclosure program is notable precisely because it builds a human-reviewed disclosure process around AI-generated findings.

The same principle applies to independent security organizations.

AI does not change responsible disclosure.

It makes responsible disclosure more important because the volume and speed of discovery may increase.

ShabuShabu’s own Responsible Disclosure framework belongs in that same defensive-security philosophy: validate carefully, minimize unnecessary exposure and coordinate remediation rather than treating vulnerability publication as the primary objective.

What Mythos-Class Models Could Change in Penetration Testing

Penetration testing will probably become increasingly hybrid.

AI handles more attack-surface analysis.

AI reads larger amounts of code.

AI helps compare roles and workflows.

AI generates vulnerability hypotheses.

AI helps explore related attack paths.

Human researchers determine which paths are credible.

They control active testing.

They validate exploitability.

They decide when sufficient evidence exists.

They interpret business impact.

They communicate remediation.

This is not the disappearance of professional penetration testing.

It is the transfer of more repetitive analysis to increasingly capable systems.

What Frontier AI Means for Security Teams Today

Organizations do not need access to Mythos 5 specifically to learn from Project Glasswing.

The strategic lessons are already useful.

Vulnerability discovery is becoming faster.

Candidate volume will increase.

Human validation needs to become more efficient.

Coordinated disclosure needs to scale.

Remediation needs more automation.

Pre-release security becomes more valuable.

Authorization and Rules of Engagement remain essential even when AI performs part of the analysis.

These changes affect every organization building software, regardless of which model it ultimately uses.

Defensive AI Should Create an Advantage Before Attackers Receive the Same Capability

The central strategy behind Project Glasswing is essentially temporal.

Powerful cybersecurity capabilities are emerging.

Those capabilities will not remain exclusive indefinitely.

Defenders therefore need to use the period before broad proliferation to find and remove vulnerabilities that future attackers might otherwise discover.

Anthropic explicitly describes Glasswing as an attempt to help defensive organizations adapt before similarly capable cyber models become widely available.

That is a sensible security objective even outside the Mythos ecosystem.

Find weaknesses while they are still defensive findings.

Patch them before they become attacker knowledge.

The Future of Defensive Cybersecurity Is Faster, but It Still Needs Judgment

Claude Mythos and Project Glasswing demonstrate that AI is moving from assisting ordinary coding toward performing increasingly sophisticated security research.

Models can inspect large codebases.

They can generate candidate vulnerabilities.

They can contribute to penetration testing.

They can assist with patch development.

They can support pre-release security reviews.

The result is a major increase in potential defensive capacity.

But increased capacity does not eliminate the core requirements of cybersecurity.

Authorization still matters.

Human validation still matters.

Severity still depends on real impact.

Coordinated disclosure still matters.

Patching still matters.

Retesting still matters.

The strongest future security programs will therefore not ask whether AI should replace security researchers.

They will ask how security researchers can use frontier AI to investigate more software while preserving the validation standards that make vulnerability research trustworthy.

That is the real promise of Mythos-class cybersecurity.

AI can make vulnerability discovery dramatically faster. Defensive security succeeds only when verification and remediation become faster with it.

Frequently Asked Questions About Anthropic Mythos and Project Glasswing

What is Claude Mythos?

Claude Mythos is Anthropic’s model line for unusually capable frontier-model use cases, including cybersecurity. Claude Mythos Preview was introduced in April 2026 with strong demonstrated computer-security capabilities, followed by Claude Mythos 5 in June 2026.

What is Project Glasswing?

Project Glasswing is an Anthropic initiative launched in April 2026 to apply advanced AI cybersecurity capabilities to securing important software. Participants have used Mythos models for vulnerability discovery and other defensive-security work.

Is Claude Mythos 5 publicly available to everyone?

No. As of August 19, 2026, Anthropic describes Mythos 5 as available through limited trusted-access arrangements rather than unrestricted general access. Access was restored to an approved set of U.S. organizations on July 1 after a temporary June suspension.

Can Mythos find software vulnerabilities?

Anthropic reports that Mythos Preview has been used to identify large numbers of vulnerability candidates across real software, with human researchers subsequently triaging and validating findings before coordinated disclosure.

Does Anthropic automatically report every vulnerability found by Mythos?

No. Anthropic’s public disclosure documentation describes a human-triage process in which candidate findings are independently reviewed before high- or critical-severity vulnerabilities are disclosed to maintainers.

Can Mythos help fix vulnerabilities?

Anthropic reports that Glasswing participants have used Mythos Preview to assist with patches and pre-release security checks in addition to vulnerability discovery.

Why is Mythos access restricted?

Anthropic says Mythos 5 has unusually powerful cybersecurity capabilities that could provide substantial defensive value but could also be misused for cyberattacks. The company therefore provides access through trusted programs rather than unrestricted general availability.

Will frontier AI replace penetration testers?

Frontier models can increasingly automate code analysis, vulnerability hypothesis generation and other research tasks, but real penetration testing still requires authorization, runtime validation, business context, production-safety judgment and human verification of security impact. Project Glasswing itself retains human triage around AI-generated findings.

What is the biggest challenge created by AI vulnerability discovery?

One emerging challenge is processing findings quickly enough. Anthropic’s Glasswing updates say the bottleneck is increasingly shifting from finding vulnerabilities toward verifying, disclosing and patching them.