Claude Mythos Preview comes to Vertex AI for a limited time and boosts cybersecurity testing

Anthropic has put Claude Mythos Preview in limited preview. The release is now available on Vertex AI from Google Cloud to a select group of users. The model stands out for its performance in cybersecurity tasks.
The announcement took place this week. Anthropic chose not to release the model widely due to its potential for use in attacks. Instead, the company created Project Glasswing, an initiative that brings together large technology companies to apply the model to system defense.
Detalhes from launch on Vertex AI
The Claude Mythos Preview preview appears on the Google Cloud blog. Selected Clientes from Vertex AI gain controlled access. The platform allows partners to test the model’s capabilities in real-world code analysis scenarios.
Anthropic published a technical report on April 7, 2026. Nele, researchers describe how the model identifies vulnerabilities in open source and proprietary projects. The focus is on high-severity faults that remain unpatched in many cases.
The model does not rely solely on known patterns. Ele analyzes code structure, understands developers’ intent, and combines logical flaws to form exploitation paths. Essa approach differs from traditional fuzzing tools, which send random data in large volume.
- Claude Mythos Preview discovered thousands of high and critical severity vulnerabilities.
- Mais of 99% of them have not yet received fixes, according to the report.
- The model autonomously created working exploits across multiple tests.
- Human Especialistas validated 89% of severity ratings with exact agreement.
Casos specific vulnerabilities found
One example involves OpenBSD. The model located a 27-year-old flaw in TCP SACK processing. The vulnerability allowed a remote denial of service attack via integer overflow in sequence number comparison.
Outra discovery occurred in FFmpeg. A 16-year-old flaw in the H.264 codec allowed writes outside memory limits. The problem arose under specific slice number conditions that traditional fuzzing tools did not often reproduce.
On FreeBSD, the model identified CVE-2026-4747. Trata discovered a 17-year-old vulnerability in the NFS server that allowed remote code execution and escalation to root privileges without authentication. The constructed exploit divided actions into multiple RPC requests.
Esses findings appear in the Anthropic technical report. Patches have already been applied in some cases, such as OpenBSD and FFmpeg. The company follows a coordinated disclosure process with deadlines for maintainers to correct flaws before full details become public.

Comparação with previous models
The performance of the Claude Mythos Preview surpasses previous versions of the Claude family. In internal tests, the model showed a significant leap in building exploits against mechanisms such as Firefox’s JavaScript. Relatórios indicate a significant increase in the number of successful exploits.
Anthropic has not released the model for general use. Access occurs through partnerships on Project Glasswing. Empresas like Amazon, Apple, Microsoft and Google participate in the initiative to strengthen the security of critical software and open source projects.
Google Cloud mentions integration into Vertex AI as a way to offer frontier models to qualified customers. Availability is restricted and is intended for defensive use.
What changes for developers and security teams
Desenvolvedores can use similar tools to review code in greater depth. The model helps identify discrepancies between what the code is supposed to do and what it actually implements. Essa capability speeds fixing problems that humans or automatic scanners are slow to find.
Anthropic reinforces its commitment to responsible disclosure. The company works with partners to report vulnerabilities and allow time for fixes. The report highlights that the biggest benefit is on the defensive side.
Equipes security now features a tool that processes large code bases quickly. The process includes logical analysis and construction of attack scenarios. Resultados undergo human validation before taking concrete actions.
Limited-time release on Vertex AI allows for controlled testing. Authorized Usuários explore functions in supervised environments. Anthropic plans to use the learnings to guide the development of future models with appropriate safeguards.
Ficha technical and current availability
Preview is active in Vertex AI for the selected group. Full Detalhes access requirements appear in the Google Cloud documentation and on the Anthropic channels.
The model fits into advanced cyber capability assessments. Benchmarks show high performance in tasks that involve reading source code, reasoning about failures and generating proofs of concept.
The initiative reflects the balance that AI companies seek. Powerful Capacidades receive restrictions to reduce risks of abuse while maximizing the potential to improve collective security.
Anthropic continues to monitor usage within Project Glasswing. Parcerias include donations and credits to support open source projects. The focus remains on protecting critical infrastructure.






