Anthropic Glasswing Changes Balance in Cyber Security

One month. More than 10,000 high and critical vulnerability. According to the update that Anthropic published on May 22, 2026, Project Glasswing reached this number in the world’s most critical software with the unpublished Claude Mythos Preview model. The figure not surprisingly, the direction change of balance: no longer find open, it has become a bottleneck to verify and patch the ones they are located. The functioning of Glasswing in this article addresses the performance, partner results and meaning for Turkey.

What is Project Glasswing And What Does It Purpose?

Project Glasswing is a collaborative cybersecurity initiative that Anthropic launched in April 2026. Basic idea is simple: before the new generation of artificial intelligence models are used for attack purposes, to secure the world’s critical software with the same models. The initiative focuses on organizations that develop and maintain software on internet and basic infrastructure.

Partner list consists of heavy balls of the sector. According to the news published in Fortune Companies like Amazon Web Services, Apple, Broadcom, Cisco, Crowd Services, Google, JPMorganChase, Microsoft and Nvidia have access to Mythos Preview for security works for defense purposes. In addition to these, approximately 40 companies that develop critical software infrastructure can access models to scan their systems and open source codes.

Anthropic undertakes 100 million USD model usage credit to venture. The study conducted with approximately 50 partners between the names such as Linux Foundation and Palo Alto Networks is supported by donations for open source ecosystem. The company’s goal is to use the model in a controlled and defense-oriented environment.

Anthropic does not open the public. The software explains new according to the rule that the industry has adopted for a long time, announced 90 days after discovery to update end users. Therefore, the company offers examples and aggregate statistics on the performance of the model instead of sharing all the details of the findings of its partners. More technical details will be published after the spread of the patches widely.

Claude How Does Mythos Preview Work?

Claude Mythos Preview is the most advanced frontier model, which Anthropic does not open to public yet. The distinctive feature of the model can detect the software open as autonomous and produce exploit codes running for this explains. So it doesn’t just say “there is a poorness in the book”; it also shows how pneumonia can be exploited. This ability points a fundamental difference from traditional scan tools.

The power of the model comes from strong agencyal coding and reasoning skills. You can analyze complex code bases, double files can be subjected to black box testing and perform leak tests. At work under Glasswing, local insufficiency detection is focused on tasks such as endpoint safety and penetration test of systems.

So why is a strong model hidden? The concern of Anthropic is clear. There is a chance to turn the same capabilities into large-scale, artificial intelligence-supported cyber attacks in malicious hands. The company states that Mythos Preview is not planning to open to general use, it is limited to a closed consortium until it has been developed sufficiently strong security measures. Part of the risks in the background of this approach, harmful artificial intelligence models Home overlaps directly with the previously addressed threats scenarios.

How Much Forward From Mythos Preview Previous Models?

The leap of Mythos Preview is understood with the best comparative data. The previous frontier model of Anthropic Claude Opus 4.6 failed in autonomous exploit development tasks. Mythos Preview performs on all models available in independent tests.

Independent assessments support this table. The UK’s AI Security Institute reported that Mythos Preview’s two “cyber range” tests comprised of very stepd cyber attack simulations. Independent security platform XBOW defines the model as a marked leap according to all models available in comparison of web exploit. ExploitBench and ExploitGym also points Mythos Preview as the most powerful model.

This difference shows that AI jumps in cybersecurity. Jeff Williams, the founder of OWASP and the CTO of Contrast Security, CSO Online analysis The situation summarizes.

"Mythos clears the first stone: When the frontier artificial intelligence starts on a large-scale open hunting, the logic of paying people for routine discovery begins to collapse. "

Jeff Williams, Founder of OWASP, Contrast Security

What Results did Partners Get?

The concrete outputs of the first month show that the model is much more future than the laboratory promise. Most of the partners have found hundreds of high or critical importance in their own software. Many reported that the open finding speed increased more than 10 floors.

Cloudflare sample is particularly remarkable. Cloudflare’s data shared on his blog The company found 2,000 bugs in critical road systems; their 400 was high or critical. Moreover, the wrong positive rate of the model was better than human test experts according to the Cloudflare team. This means both speed and accuracy.

Data from other partners also shows the same direction:

Mozilla: Firefox found 27 open in 150 and patched. This is the old version of Firefox with Claude Opus 4.6 in 148.

Palo Alto Networks: In the latest version, the usual 5 solid patch released.

Oracle: Clears in products and cloud are quickly finding and corrected from the previous floor.

Microsoft: It reported that the number of new patches to release will continue to increase for some time.

The model also worked beyond the classic poor screening. As transferred in Engadget Mythos Preview on a bank of Glasswing helped detect and block 1.5 million dollars of fake weather made after a customer’s e-mail account has been addressed.

Easy to Find On, Why Is It Hard to Patch?

The most important bull of Glasswing is not a figure, but a reversal. The official update of Anthropic As mentioned, the progress in software security was limited to how quickly it could be found. Now the limit is how quickly we can verify, report and patch a large number of deficits found by artificial intelligence.

The numbers show this bottle clear. Anthropic detected 23.019 problems by scanning more than 1,000 open source projects; they predicted 6.202 as high or critical value grade. Six independent security firms studied 1.752 from these findings and confirmed 90%,6 in real positive. If this verification rate lasts, even if the model does not find any other open, it will only appear about 3.900 high or critical open in open source code. But only 75 patched from the reported 530 high/critic open, the public warning for 65 was released.

A concrete example that Mythos Preview finds is that the cliff appears. The model detected an open source cryptography library used in billions of devices at wolfSSL and produced a exploit that works for it. This exploit could allow a attacker to host a completely legitimate fake site in appearance of a bank or email provider by producing fake certificates. Open patched with CVE-2026-5194 number.

This is a reason for slowness, the capacity of caregivers that volunteer open source projects. Some caregiers asked Anthropic to slow down the description speed, because they need more time to design patches. A high or critical open with Mythos Preview is patched in the average two weeks. A size of the problem is that the carekeepers are already overwhelming with poor bug reports, low quality, artificial intelligence production. Anthropic has partnered with the Alpha-Omega project of Open Source Security Foundation to alleviate this load and aims to increase the processing capacity of maintenancers.

Practical advice for defenders is not shown but effective. Software developers should shorten patch loops and make updates easy for users. Network defenders focus on basic controls that institutions such as NIST and UK NCSC: multi-factor authentication, tightened default configurations and detailed record retention. Anthropic provides code base mapping tools and automatic threat model makers for partners; Cisco has also made its own Foundry Security Spec source open source.

What Does Understanding Come For Turkey?

While Glasswing is directly executed with a partner from Turkey, its effects are extending to the local ecosystem. Most of the described open-source libraries that make the basis of the internet. These libraries work under many domestic systems from banks in Turkey, e-commerce platforms to mobile applications. The patching of a wolfSSL or Firefox deficit is interested in anyone using the chain.

The main message for security teams in Turkey, Mythos class models will soon open to wider access. Anthropic also states that models with similar capabilities will be developed by other companies in a short time. This means that the same tools are accessible for both defenders and attackers. Teams that shorten patch loops from now, implement basic controls will be advantageous during this transition period.

On the practical side, even public models can find a large number of outdoors. Anthropic released Claude Security tool for Claude Enterprise customers as open beta to everyone, this tool helped to patch more than 2,100 corporate vulnerability in three weeks with Claude Opus 4.7. Some of the tools available for Turkish developers who want to scan vulnerability on code bases, previously discussed AI code writing sites Home between. Offering similar security deficit detection free artificial intelligence coding toolscan be a good starting point for small teams.

What does the community think?

Supporters

Supporting the initiative defends that the defenders have earned an asymmetric advantage against attackers for the first time. Finding and closing the openness of the critical infrastructure before the attackers reduces the risk of the billions of end users. The model of Anthropic is seen as a responsible approach to the limitation of a closed consortium rather than open general use. The support provided to donations and carekeepers made in open source ecosystem is among steps that strengthen the entire ecosystem.

News

The criticists first point to the bottle. The less than 1% of the open spaces found may not be patched, but may not work other than adding loads to an already overloaded security ecosystem. Regarding the second criticism timing. Analysis of Picus Security As mentioned, Glasswing was announced when Anthropic reached an important income mileage stone and a major deal of processing power; while some commentators find useful for the project sector, but also read as a strong marketing move for Claude.

Frequently Asked Questions

What is Project Glasswing?

Project Glasswing is a collaborative cybersecurity initiative that Anthropic launched in April 2026. Approximately 50 large technologies and security companies aim to find and close open in critical software using the unpublished Claude Mythos Preview model.

Claude Is Mythos Preview open to public?

No. Mythos Preview is currently open to a closed partner consortium. Anthropic states that the model does not plan to open public use until it is developed sufficiently strong security measures to prevent abuse.

How many open mythos Preview per month?

Partners together found more than 10,000 high or critical importance ratings. Also 23.019 problems were detected in the open source scan of Anthropic; their 6.202 were highly predicted or critical.

How different is Mythos Preview from previous models?

Previous frontier model Claude Opus 4.6 failed to develop autonomous exploit. Mythos Preview performs both this task and performance on all models available in independent benchmarks.

Are the developers in Turkey access to these tools?

Although access to Mythos Preview is limited, Anthropic released Claude Security tool for Enterprise customers as beta open to everyone. Public models can also already find a large number of open models.

Anthropic plans to expand Glasswing with new partners including US and allied governments. Mythos class models are open to general use, only after being activated more powerful security measures. This transition period, which is still expensive and slow if it is cheap, patching, will greatly determine the future of critical software in the coming months.