Anthropic’s Unreleased Claude Mythos Might Be The Most Advanced AI Model Yet
Anthropic is testing an unreleased artificial intelligence (AI) model with capabilities that exceed any system it has previously released, and an accidental data leak forced the company to confirm what it had not yet announced.
According to a Thursday (March 26) report by Fortune, a configuration error in Anthropic’s content management system left nearly 3,000 unpublished documents in a publicly searchable data store, including a draft blog post describing the model as “by far the most powerful AI model we’ve ever developed.” Independent security researchers at LayerX Security and the University of Cambridge found the materials before Anthropic restricted access.
An Anthropic spokesperson confirmed the model’s existence to Fortune, calling it a step change and the most capable system the company has built, with meaningful advances in reasoning, coding and cybersecurity.
The company said it is testing the model, known as Claude Mythos, with a small group of early-access customers and has not set a general release date, partly because it remains expensive to run at scale.
What Mythos Can Do
The leaked draft described Mythos as part of a new model tier called Capybara, positioned above its current top-tier Opus models in both capability and cost. Where prior models respond to instructions one step at a time, Mythos plans and executes sequences of actions on its own, moving across systems, making decisions and completing operations without waiting for human input at each stage.
As reported by Fortune, the leaked document described Mythos as currently far ahead of any other AI model in cybersecurity capabilities and said it signals an approaching generation of systems that can find and exploit software weaknesses faster than defenders can close them. Anthropic said its rollout plan prioritizes enterprise security teams, giving defenders early access before the model reaches wider distribution.
According to a Sunday (March 29) report from Axios, Anthropic has been privately warning 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 carry out complex operations with minimal human involvement.
The Attack Record That Preceded It
According to a separate Axios report, published in November, a Chinese state-sponsored hacking group in September used an earlier Claude model to carry out 80-90% of a coordinated attack campaign on its own, working through roughly 30 organizations including technology companies, financial institutions and government agencies before Anthropic detected and shut it down.
The AI identified targets, found weaknesses, wrote attack code and produced detailed post-operation reports, all with minimal human direction. The operators running the attack convinced the model it was performing legitimate security testing. Once inside that framing, the AI executed the operation without further instruction.
A Dark Reading poll published in January found that 48% of cybersecurity professionals now rank agentic AI as the top attack vector for 2026, above deepfakes and social engineering.
What the Industry Has Already Documented
As reported by PYMNTS, the September Claude Code incident marked the first confirmed case in which an AI agent handled most steps of a cyberattack normally performed by human operators.
Eva Nahari, then-chief product officer at AI solutions firm Vectara, told PYMNTS the campaign was “global, industry-agnostic and growing,” adding that with automation comes velocity and scale, and that attackers are now acquiring the same advantages that AI gives enterprises.
As also reported by PYMNTS, Anthropic’s earlier research found that its Claude Opus 4.5 model reduced successful prompt injection attacks to 1% in browser-based operations, down from higher breach rates in earlier versions, though the underlying vulnerability persists as browser-based automation grows more common. PYMNTS Intelligence found that 98% of business leaders remain unwilling to grant AI agents action-level access to core systems, with trust as the primary constraint on adoption.
According to a Monday (March 30) report by CSO Online, shares of major cybersecurity vendors, including CrowdStrike, Palo Alto Networks, Zscaler and Fortinet, fell following the Mythos news as investors considered what frontier AI capabilities embedded in security tools could mean for the industry’s competitive structure.
The post Anthropic’s Unreleased Claude Mythos Might Be The Most Advanced AI Model Yet appeared first on PYMNTS.com.