Anthropic CEO Dario Amodei said the company sacrificed a multi-month competitive lead by holding back broader access to its Mythos-class AI models from February to June 2026 to create stronger cybersecurity protections. Amodei laid out that reasoning in a lengthy post on X published Tuesday, responding to criticism from AI safety advocate Nathan Calvin and defending the firm's approach to managing risks as models grow more powerful. The disclosure came as part of a broader debate over whether Anthropic is weakening its safety commitments while racing to build superintelligence.

Anthropic first unveiled Claude Mythos Preview in April through Project Glasswing, granting selected partners including Amazon, Apple, Microsoft, CrowdStrike, and the Linux Foundation access to the model's vulnerability-discovery tools for defensive purposes while keeping it away from ordinary developers. The company expanded the program on June 2 to roughly 150 more organizations across more than 15 countries. It released Claude Fable 5 on June 9, a Mythos-class model with added cybersecurity and biological safeguards suitable for general use, but suspended both Fable 5 and its sibling Mythos 5 three days later after a U.S. government export-control order barred access by foreign nationals, including Anthropic's own staff. The Commerce Department lifted those controls on June 30, and Fable 5 returned worldwide on July 1 while Mythos 5 had already regained access for select U.S. organizations on June 26.

According to Amodei, releasing the models sooner would have given organizations less time to prepare their infrastructure for increasingly capable AI-powered cyberattacks. Anthropic tells The New Stack that the delay follows a commitment in its Responsible Scaling Policy to require a strong argument that catastrophic risk is contained and to delay development and deployment as needed until it no longer believes it holds a significant lead. The company warns that for most partners, a major attack could affect more than 100 million people by its estimate, and says "hundreds of thousands of organizations, researchers, and maintainers" will likely need advanced cybersecurity capabilities before it can release Mythos-level models more broadly. Calvin, the critic who prompted Amodei's response, questions why Anthropic's mistakes "systematically tend to guard and favor Anthropic's interests," pointing to the firm's more than $500 billion in computing deals and changes to its Responsible Scaling Policy that removed a 2023 pledge never to train an AI system unless the company could guarantee its safety measures were adequate in advance.

Amodei conceded that AI development could soon outpace the safeguards built to govern it, writing "we are nearing that point." The company's own research illustrates the concern: as of May, Claude authored more than 80% of the code merged into Anthropic's codebase, a figure the company calls conservative and leadership has separately estimated at 90% or more when scripts and experimental code are included. That statistic measures lines of production code attributed to Claude, not autonomous control over engineering decisions, and the authors write that "recursive self-improvement is not inevitable." Amodei defended the company's approach as a "race to the top," arguing that Anthropic can shape industry safety standards more effectively by staying at the technological frontier, and pointed to restrictive biological safeguards, support for AI regulation, and export controls on advanced chips as evidence the firm puts safety ahead of short-term profits even when those decisions carry substantial costs.

Anthropic's Frontier Safety Roadmap sets December 15, 2026, as the target for a policy proposal on how governments can support deeper access for embedded evaluators, and by January 15, 2027, the company plans to either publish a policy framework to shape the pace of AI development or explain why it decided not to. The roadmap warns that automated AI research "could accelerate progress faster than governance or safety measures can keep up," and the company acknowledges its roadmaps may change, saying it will publish updates on whether it meets its goals. Amodei outlined a three-stage plan in September that begins with independent evaluators working inside AI companies, expands to coordination among democratic countries, and ultimately calls for agreements involving both democratic and authoritarian governments, arguing that slowing the frontier requires industry-wide and global coordination alongside steps individual labs can take alone. The tension between maintaining commercial competitiveness and honoring safety pledges will likely define how the industry navigates the next phase of capability growth. Whether voluntary commitments can survive market pressure remains an open question that extends well beyond any single company's intentions.