Anthropic’s latest artificial intelligence model, Claude Oceanus-v1-p, has entered a testing phase under controlled conditions. However, the model’s early distribution was compromised even before the formal assessment began, raising significant concerns.
Unauthorized Distribution Raises Concerns
On June 3, 2026, researchers began noticing references to the model claude-oceanus-v1-p within Anthropic’s Claude Console and through unauthorized API proxy services. This discovery led to speculation about Anthropic’s plans for a wider release of a successor to their Claude Mythos line.
The model was quickly accessed by red team evaluators, but the testing period was abruptly interrupted. Reports soon emerged that an unidentified entity had allegedly resold API access to the model through a proxy service based in China. The resale was at a premium rate of $16 per million input tokens, considerably higher than Anthropic’s usual pricing.
History of Proxy Abuse
Anthropic has a well-documented history with unauthorized proxy abuse incidents. Earlier in 2026, the company accused several Chinese AI labs, including DeepSeek, Moonshot AI, and MiniMax, of using tens of thousands of fake accounts to engage with Claude models via proxy channels.
In response to the latest breach involving Oceanus, Anthropic temporarily halted model access for the wider red team group while conducting an internal investigation into the matter.
Implications for AI Cybersecurity
Claude Oceanus-v1-p builds upon the Claude Mythos Preview foundation, introduced in April 2026, which displayed significant capabilities that intrigued the cybersecurity community. The Mythos Preview, under Anthropic’s restricted research track, was found capable of identifying and exploiting zero-day vulnerabilities across major operating systems and browsers.
The Frontier Red Team and Glasswing partners discovered over 10,000 high or critical-severity vulnerabilities, with the Turing Institute noting a recovery rate exceeding 99% in disclosed test cases.
In tandem with the Oceanus evaluation, Anthropic expanded its Project Glasswing, an AI cyberdefense initiative, to include around 150 new organizations across more than 15 countries. This expansion now covers vital sectors such as power, water, healthcare, and communications, previously not included when the program initially focused on Big Tech.
Anthropic acknowledges that a significant cyberattack on these new partner organizations could potentially impact over 100 million people. The company has emphasized that neither Mythos-level capabilities nor the Oceanus-v1-p model will be publicly available until robust safeguards are developed to prevent misuse, a standard not yet met by the industry.
