Anthropic Confirms Leaked Model Marks 'Step Change' in Reasoning After Data Breach

A security misconfiguration at Anthropic has exposed internal documents revealing details of an unreleased AI model that internal teams describe as the company's most capable to date. After Fortune reported the breach, Anthropic confirmed it is already testing the model with select customers, characterizing it as marking a "step change" in reasoning, coding, and cybersecurity capabilities.

How the Breach Occurred

The exposure resulted from a misconfiguration in Anthropic's content management system. A default setting automatically made all uploaded files public, leaving approximately 3,000 internal documents accessible to the internet without authentication or authorization controls.

The specific technical details about the unreleased model remain limited, as Anthropic has not officially announced the system or provided performance metrics, parameter counts, or availability timelines.

Broader Context: OpenAI Also Preparing Major Release

Anthropics's unreleased model announcement comes as OpenAI reportedly prepares its own major capability jump. OpenAI is developing a model internally codenamed "Spud," which has completed pretraining. CEO Sam Altman has internally stated the model can "really accelerate the economy," though OpenAI has not officially disclosed specifics about the system's architecture, capabilities, or release date.

Strategic Timing and IPO Implications

Both companies may be timing their strongest model releases to align with planned IPO activity later in 2026. This suggests major announcements could arrive within the coming months as both organizations position themselves for public markets.

Neither Anthropic nor OpenAI has provided confirmed details about pricing, context window size, benchmark scores, or other technical specifications for their respective unreleased models.

What This Means

The breach underscores persistent security challenges in AI infrastructure, even at well-resourced organizations. More significantly, both Anthropic and OpenAI are signaling that substantial capability improvements are imminent—though neither company has yet demonstrated these advances publicly. The gap between internal testing and public release typically involves safety evaluation, red-teaming, and regulatory assessment, meaning months may pass before customers can access either system. The coincidence of parallel releases from both organizations suggests an intensifying arms race in AI capability development heading into 2026.