Anthropic’s latest AI model, Claude Oceanus-v1-p, has surfaced in restricted testing channels, but its early distribution was compromised before formal evaluations began. References to ‘claude-oceanus-v1-p’ appeared on June 3, 2026, within Anthropic’s Claude Console and unauthorized API proxy services, leading to speculation about an impending successor to the Claude Mythos line.
Shortly after red team evaluators gained access, reports emerged of an unidentified actor reselling API access to ‘claude-oceanus-v1-p’ through a Chinese-based proxy service at $16 per million input tokens, significantly above Anthropic’s standard enterprise rates. This incident echoes earlier in 2026 when Anthropic accused Chinese AI labs of using approximately 24,000 fake accounts to run over 16 million interactions with Claude models via proxy channels.
In response, Anthropic paused model access for the broader red team cohort pending an internal investigation. Claude Oceanus-v1-p builds upon the Claude Mythos Preview foundation, which launched in April 2026 and demonstrated capabilities such as identifying and exploiting zero-day vulnerabilities across major operating systems and web browsers. The Turing Institute noted that Mythos’ red team found vulnerabilities with a recovery rate exceeding 99% across disclosed test cases.
This development follows Anthropic’s June 2 expansion of Project Glasswing, its restricted AI cyberdefense initiative, to approximately 150 new organizations across more than 15 countries, including India, France, Germany, South Korea, and Australia. The expanded group now includes critical infrastructure sectors like power, water, healthcare, and communications, with potential cyberattacks affecting over 100 million people.
Anthropic has stated that Mythos-level capabilities, and by extension, Oceanus-v1-p, will not be cleared for general public release until the company develops ‘highly robust safeguards to prevent misuse,’ acknowledging that such safeguards do not yet exist in the industry.
The premature exposure of Claude Oceanus-v1-p underscores the challenges in securing advanced AI models during development. As AI capabilities advance, ensuring controlled distribution and robust safeguards becomes increasingly critical to prevent misuse and maintain trust in AI technologies.
Source: CyberSecurityNews
