Anthropic releases preview of new AI model Mythos for cyber defence

Anthropic, a prominent player in the artificial intelligence sector, has introduced a glimpse of its innovative AI model, Mythos, aimed at enhancing cybersecurity measures. This cutting-edge model is set to play a crucial role in the company’s ambitious initiative, “Project Glasswing,” which unites leading technology firms to leverage AI for bolstering defensive security. According to a recent blog post from Anthropic, Mythos is poised to identify vulnerabilities within essential software systems. The company highlighted that the Mythos Preview has already detected thousands of high-severity vulnerabilities, impacting all major operating systems and web browsers. So, what exactly is Project Glasswing? This initiative brings together a coalition of tech giants, including Amazon Web Services (AWS), Apple, Google, Microsoft, and many others, alongside over 40 additional organizations. Their collective aim is to safeguard critical software infrastructure by utilizing the capabilities of Mythos Preview, ensuring that both proprietary and open-source systems are fortified against threats. Anthropic emphasized the necessity of collaboration, stating, “No one organization can tackle these cybersecurity challenges alone.” The initiative is driven by the capabilities of Claude Mythos Preview, designed to decode intricate data and patterns, effectively outperforming most human analysts in identifying software vulnerabilities. The features of Mythos Preview are impressive, with capabilities for detecting previously unknown zero-day vulnerabilities and supporting key cybersecurity workflows such as penetration testing and binary analysis. Its integration into both enterprise and open-source security processes enhances endpoint and infrastructure security, while automating vulnerability detection and triage processes. Noteworthy achievements include the identification of a 27-year-old vulnerability in OpenBSD and a 16-year-old flaw in FFmpeg, showcasing Mythos’ robust performance in vulnerability detection. Performance benchmarks reveal that Mythos Preview achieved an impressive 83.1% on the CyberGym benchmark for vulnerability reproduction, significantly outperforming its competitor, Opus 4.6. To further support participants in Project Glasswing, Anthropic is offering up to $100 million in model usage credits. The model will be accessible through various platforms, including the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry, although public access remains limited to select partners. Security professionals can request access through the Cyber Verification Program and Claude for the Open Source initiative.