Anthropic to Partially Release Highly Dangerous AI "Claude Mythos"

Anthropic has expanded access to its advanced AI model “Claude Mythos,” which was previously deemed too dangerous for public release. According to a report by The Conversation, this specialized AI model is designed to automatically detect software vulnerabilities. As part of an initiative known as Project Glasswing, Claude Mythos is being made available to about 150 organizations across 15 countries, including the Australian government. While it is expected to give defenders a crucial advantage in the realm of cybersecurity, concerns have been raised about the risks of misuse due to the model’s powerful capabilities.

Too Dangerous for Public Release

Claude Mythos is distinctly different from traditional general-purpose large language models. It is explicitly designed to analyze code and pinpoint critical vulnerabilities that might be overlooked by humans. Unlike typical AI models geared toward tasks like chat or text generation, this model focuses solely on identifying bugs and security loopholes.

Anthropic has hailed the model as having “the potential to deliver a definitive victory to defenders,” but it has also imposed strict access control due to concerns about the potential damage if such capabilities were to fall into the hands of malicious actors. Although its release remains limited, this expansion marks a turning point by providing top government agencies and major corporations with state-of-the-art AI defenses.

Track Record of Detected Vulnerabilities

and Accuracy

In initial testing, Claude Mythos flagged 23,000 vulnerabilities, with an estimated 6,200 categorized as high-risk. Human experts later confirmed that approximately two-thirds of these high-risk vulnerabilities were indeed severe.

While the original report did not provide specific comparative metrics, the claim that “defenders can, for the first time, gain a decisive advantage over attackers” suggests unprecedented performance. Particularly noteworthy is the model’s improved ability to identify zero-day vulnerabilities, which could significantly enhance the security infrastructure of both corporations and governments.

Risks of False Positives and Overload

However, the deployment of Claude Mythos is not without its challenges. Of the 23,000 flagged vulnerabilities, only around 4,000 were confirmed to be high-risk after verification. The majority were either false positives or low-risk, which could overwhelm security teams with a sheer volume of alerts.

In the current cybersecurity landscape, alert fatigue is already a pressing issue. If automated tools generate excessive noise, there is a higher risk of missing genuinely critical threats. Organizations adopting Mythos must concurrently develop strategies for prioritizing alerts and reallocating human resources effectively.

Threats Against the AI System Itself

In addition to the vulnerabilities that Claude Mythos identifies, the AI system itself could become a target for attacks. The original article cites an example involving Meta’s AI chatbot being exploited, leading to the compromise of high-profile accounts, including that of former U.S. President Barack Obama. Techniques like prompt injection and model manipulation have increasingly been used to force AI systems into unintended behaviors.

Given the high value of a system like Claude Mythos, it is likely to become a prime target for attackers. Organizations granted access to it must also prepare for potential internal threats and supply chain vulnerabilities that could jeopardize the AI’s integrity.

Australian Government’s Participation and

Future Prospects

The Australian Signals Directorate (ASD) has expressed enthusiasm about participating in Project Glasswing. For a country that has recently faced a series of major cybersecurity incidents involving companies like Optus and Medibank Private, Mythos could serve as a much-needed defensive tool. However, the government has yet to disclose detailed plans for its implementation or evaluation criteria.

It remains uncertain whether this limited rollout will eventually lead to broader availability or if the project will be curtailed based on risk assessments. Anthropic may be positioning this initiative as a case study in how society can responsibly adopt and manage “high-risk AI.”

Editorial Opinion

Short-Term Impact

In the next three to six months, it is expected that organizations participating in Project Glasswing will fully implement vulnerability scans using Claude Mythos, leading to a significant improvement in the security of software supply chains. However, the high rate of false positives may exacerbate existing shortages of cybersecurity professionals. Japanese companies and government agencies might begin exploring similar initiatives, particularly in critical sectors like finance and infrastructure.

Long-Term Outlook

Over the next one to three years, “vulnerability detection AI” could become a commercially available tool, leading to the commoditization of security diagnostics. As attackers gain access to similar technologies, the defense advantage may not last indefinitely. Instead, a new battleground of AI-versus-AI cybersecurity could emerge, reshaping competition in the security industry. This may also prompt international discussions on regulating advanced AI technologies and their export. Japan will need to focus on both technological advancements and establishing legal governance frameworks to stay ahead.

Questions from the Editorial Team

Determining who should have access to dual-use AI technologies like Claude Mythos and under what conditions is a highly complex issue. Readers are encouraged to consider the role of cutting-edge AI in their organization’s security strategies and the extent to which they would trust and rely on externally provided tools. Additionally, how do you believe the adoption of such tools will impact the roles and responsibilities of existing security engineers? We welcome your thoughts on this matter.

References

• The Conversation: Australia now has access to Anthropic’s Claude Mythos — Published on June 4, 2026