Post Content
Anthropic has said that its latest upcoming AI model outperforms every other large language model (LLM) it has released to date, suggesting a major potential leap in the race to develop cutting-edge frontier AI systems.
The unreleased model marks a “step change” in AI performance and is “the most capable we’ve built to date,” an Anthropic spokesperson was quoted as saying by Fortune. The buzz around the model comes after an internal data leak exposed its existence to cybersecurity researchers Roy Paz and Alexandre Pauwels, who discovered a draft announcement blog post containing a description of the model and other details in an unsecured and publicly searchable data cache.
The leak has also been confirmed by Anthropic, which reportedly acknowledged that a “human error” in the configuration of its content management system (CMS) led to the draft blog post being accessible. While Anthropic has yet to say when it plans to unveil the new model, here’s what we know so far.
Key details about Anthropic’s next-gen AI model
The new model described in the leaked blog post is referred to as ‘Claude Mythos’. Anthropic has said that it has completed training Mythos and has opened up access to the model to a small group of early access users.
Mythos is reportedly part of an entirely new lineup of AI models by Anthropic called ‘Capybara’. Currently, Anthropic offers each of its AI models in three different sizes, namely: Opus, which is the largest and most capable model versions; Sonnet, cheaper, slightly faster but also less capable; and Haiku, its cheapest, smallest, and fastest brand of LLMs.
“‘Capybara’ is a new name for a new tier of model: larger and more intelligent than our Opus models—which were, until now, our most powerful,” Anthropic has been quoted as saying in the blog post. Models in this new tier are reportedly larger and more capable than Opus, but also more expensive to use.
Mythos’ capabilities reportedly pose unprecedented cybersecurity risks, which has Anthropic concerned about its real-world implications. “Compared to our previous best model, Claude Opus 4.6, Capybara gets dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity, among others,” the company was quoted as saying in the leaked material.
Story continues below this ad
Also Read | Anthropic’s Claude can now use your computer like a human: Will it replace OpenClaw?
“In preparing to release Claude Capybara, we want to act with extra caution and understand the risks it poses—even beyond what we learn in our own testing. In particular, we want to understand the model’s potential near-term risks in the realm of cybersecurity—and share the results to help cyber defenders prepare,” the document read.
One of Anthropic’s concerns is that the model could be misused by threat actors to run large-scale cyber attacks. Mythos is “currently far ahead of any other AI model in cyber capabilities,” and “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders,” the company notes in the blog post.
These concerns could explain Anthropic’s cautious rollout strategy of Mythos, starting with a small group of early-access users. The model is also expensive to run and not yet ready for general release, as per the draft blog post. When it is released more widely, Anthropic is expected to make it first accessible to cyber defenders. “We’re releasing it in early access to organizations, giving them a head start in improving the robustness of their codebases against the impending wave of AI-driven exploits,” it was quoted as saying.
AI disruption of cybersecurity industry
Even before its official release, Mythos’ reportedly advanced cyber capabilities sparked a sell-off of cybersecurity stocks on Friday, March 27, according to a report by CNBC. Shares of iShares Cybersecurity ETF fell 4.5 per cent, whereas CrowdStrike, Palo Alto Networks, Zscaler, and SentinelOne shares declined by 6 per cent each. Tenable stock saw the sharpest fall of 9 per cent.
Story continues below this ad
Also Read | Why OpenAI is shutting down its viral AI video app Sora
Cybersecurity companies are under immense pressure to evolve along with a shifting threat landscape spurred by the rise of AI-powered tools and autonomous agents, which have significantly lowered the entry barrier to carry out sophisticated cyber attacks. In November 2025, Anthropic said that a state-sponsored group in China used its Claude AI model to automate a cyber attack. Reports of Anthropic testing Mythos also come days after The Information reported that archrival OpenAI has finished pre-training a new, very strong model referred to as ‘Spud’.
