- Anthropic is testing an AI model internally called Mythos (also referred to as Capybara), described as “by far the most powerful AI model we’ve ever developed.”
- The model dramatically outscores Claude Opus 4.6 on software coding, academic reasoning, and cybersecurity benchmarks, representing a new tier above the Opus model family.
- Anthropic has privately warned senior government officials that Mythos makes large-scale cyberattacks “significantly more likely” in 2026.
- The model’s existence was revealed through a data leak involving an unsecured content management system before Anthropic intended to announce it.
What Happened
Anthropic is testing an AI model internally designated Mythos — also known as Claude Capybara — that the company describes as a “step change” in AI performance, according to a Fortune exclusive published March 26, 2026. Reporter Beatrice Nolan obtained details after a data leak exposed a draft blog post describing the model as “by far the most powerful AI model we’ve ever developed.”
An Anthropic spokesperson confirmed the model’s existence and described it as “the most capable we’ve built to date.” The company attributed the data exposure to “human error” in content management system configuration and removed public access after Fortune’s notification. Days later, a second security incident exposed source code for Claude Code, Anthropic’s CLI tool, raising questions about the company’s internal data handling practices during a period of rapid growth.
Why It Matters
Mythos represents a new model tier above Anthropic’s existing Opus family. Claude Opus 4.6, released in February 2026, already leads on SWE-bench Verified at 80.8% and scored 40.0% on Humanity’s Last Exam. If Mythos scores “dramatically higher” than Opus 4.6 on coding, reasoning, and cybersecurity benchmarks — as internal documents indicate — it would widen Anthropic’s lead at the frontier of AI capability.
The timing places Mythos in direct competition with OpenAI’s GPT-5.4 and Google’s Gemini 3.1 Pro. Anthropic’s $380 billion valuation, reached after a $30 billion funding round led by Singapore’s GIC and U.S. firm Coatue, reflects investor confidence in the company’s ability to maintain this competitive position.
Technical Details
Capybara is positioned as a larger and more expensive model than Opus, suggesting significantly more parameters or a different architectural approach. Internal Anthropic documents describe the model as “currently far ahead of any other AI model in cyber capabilities,” according to Fortune’s reporting.
Roy Paz, a senior AI security researcher at LayerX Security, noted the model’s potential to enable attackers to “exploit vulnerabilities in ways that far outpace” current defensive capabilities. Alexandre Pauwels, a cybersecurity researcher at the University of Cambridge, echoed concerns about the gap between offensive AI capability and defensive readiness.
Anthropic activated AI Safety Level 3 protocols for its Claude 4 models due to increased misuse potential. Whether Mythos would trigger a higher safety level — or require entirely new protocols — has not been disclosed. The company is providing early access to cybersecurity professionals as part of its cautious release strategy.
Who’s Affected
The cybersecurity implications affect every organization with an internet-facing attack surface. Anthropic has privately warned senior government officials that Mythos makes large-scale cyberattacks “significantly more likely” in 2026, and that agents running on systems at this capability level can plan and execute complex operations with minimal human involvement. Cybersecurity stocks fell after the model’s existence became public, with CNBC reporting declines across the sector.
Nearly 200 protesters from the organization Stop the AI Race demonstrated outside Anthropic’s headquarters in March, demanding a halt to AI development over safety and existential risk concerns — a reaction that a model described as the most powerful ever would likely intensify.
What’s Next
Anthropic has not set a general release date for Mythos, partly because it remains expensive to run at scale. The model is currently available only to a small group of early-access customers. The testing phase could last months before any commercial release, and the company’s history of prioritizing safety research suggests additional evaluation periods before broad availability.
The second data leak — exposing Claude Code source code just days after the Mythos revelation — adds urgency to questions about Anthropic’s operational security during a period of explosive growth. Whether the company can maintain its safety-first reputation while managing the pressures of a $380 billion valuation and a model that its own documents describe as posing unprecedented risks remains the central tension heading into the second half of 2026.
