Why Anthropic held back Claude Mythos, its most powerful AI model yet?

The race to build more capable artificial intelligence models has largely focused on making systems that can write better, code faster and solve increasingly complex problems. But Anthropic's Claude Mythos has become one of the few AI models to make headlines for what it wasn't allowed to do.Released in April 2026, Claude Mythos was introduced as Anthropic's most advanced model. Yet unlike most frontier AI systems, it was not widely rolled out to developers or consumers. Instead, the company chose to keep the model under tightly controlled access after internal testing suggested its cybersecurity capabilities could pose significant risks if released without safeguards.What is Claude Mythos?Claude Mythos is Anthropic's flagship AI model built for advanced reasoning, software engineering and cybersecurity tasks. While previous Claude models focused on coding assistance, enterprise workflows and general-purpose AI, Mythos pushed much further into offensive cybersecurity.According to Anthropic, the model demonstrated the ability to independently identify software vulnerabilities and develop exploit paths with minimal human guidance. During internal testing, it reportedly achieved a perfect score on the Cybench cybersecurity benchmark and uncovered thousands of previously unknown software flaws, including a 17-year-old remote code execution vulnerability in FreeBSD.These findings positioned Mythos as one of the most capable AI systems developed for cybersecurity research.Why didn't Anthropic release it publicly?The company concluded that Mythos' strengths also created its biggest challenge.Unlike concerns around misinformation or harmful text generation, Anthropic's primary worry centred on the model's ability to discover and exploit vulnerabilities before organisations had an opportunity to fix them.If such capabilities became widely accessible, attackers could potentially use the model to identify weaknesses in software systems faster than defenders could patch them. Anthropic also said testing revealed behaviours that required further safety evaluation before broader deployment.Rather than proceeding with a conventional product launch, the company restricted access while strengthening safety mechanisms and studying the model's behaviour under controlled conditions.Project Glasswing: Putting Mythos to defensive useInstead of shelving the technology altogether, Anthropic launched Project Glasswing, a cybersecurity initiative aimed at using Mythos for defensive purposes.The programme brought together technology companies and security organisations, including Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks. Anthropic also committed $100 million in model usage credits to support the initiative.According to the company, participating organisations identified more than 10,000 high- and critical-severity software vulnerabilities within weeks of the programme's launch. The model was also used for automated patch generation, penetration testing, threat detection and modernising legacy software systems.Project Glasswing later expanded to around 150 organisations across more than 15 countries, including operators of critical infrastructure and cybersecurity agencies.Has Claude Mythos been released now?Anthropic has gradually moved towards controlled deployment rather than a complete public release.The company has since introduced Claude Fable 5 for general AI users and Claude Mythos 5 for approved organisations with specialised cybersecurity requirements. Access to Mythos-class capabilities remains limited and is governed through verification programmes designed for defensive use.This approach reflects Anthropic's attempt to balance innovation with security by allowing vetted organisations to benefit from the model while limiting opportunities for misuse.Why does this matter for the AI industry?Claude Mythos illustrates a broader shift in how frontier AI companies are approaching increasingly powerful models.As AI systems become capable of performing tasks with real-world security implications, developers are moving beyond the traditional "launch first, iterate later" strategy. Instead, companies are adopting staged rollouts, restricted access programmes and closer collaboration with governments and enterprise partners.The debate is no longer just about building more capable AI. It is increasingly about determining when those capabilities should be released, who should have access to them and what safeguards are needed before they reach the wider public.Also read: ₹13 crore salaries and counting: Anthropic's H-1B filings reveal the cost of winning the AI race