Anthropic's Next Mythos-Class Model Leaks During Red Team Testing

On June 3, 2026, the model identifier claude-oceanus-v1-p began circulating among security researchers after it appeared inside Anthropic's internal Claude Console. The sighting immediately triggered speculation that Anthropic was moving toward a broader rollout of a successor to its restricted Mythos model line. That speculation hardened into a security incident within a few hours: an unidentified actor was reportedly offering API access to Oceanus through a Chinese-based proxy service at $16 per million input tokens, more than three times Anthropic's standard pricing for its current flagship. Anthropic responded by suspending model access for the broader red team cohort, pending an internal investigation that is still underway.

What the Leak Reveals About Oceanus

Claude Oceanus-v1-p appears to be the next stage in Anthropic's Mythos family, the high-capability line the company has kept out of public release since April 2026, citing the risk that its advanced security capabilities could be misused before defenders can patch what it finds. The original Claude Mythos Preview demonstrated the ability to autonomously discover thousands of zero-day vulnerabilities across every major operating system and browser, working far faster than any human security team.

A partial system prompt that circulated alongside the leaked identifier points to several capability additions over Mythos Preview. The prompt references enhanced code interpreter access, direct database querying, and a "thinking budget" parameter that lets the model allocate additional compute time to complex problems on the fly. Red team evaluators who had access before the suspension described the model's focus as advanced reasoning, coding, long-horizon agentic work, and cybersecurity, the same general profile as Mythos but with broader scope and more flexible self-direction.

From Evaluation to the Open Market

Anthropic's red team programs for its Mythos-class models run through a dedicated subdomain and are restricted to organizations in the Project Glasswing coalition, the roughly 150 critical infrastructure partners that were given Mythos Preview access over the past two months. How the Oceanus identifier moved from that environment into unauthorized circulation has not been explained publicly.

The reported timing matters. Red team programs at Anthropic typically run about seven days before any broader launch, which would put Oceanus at a very early stage of external evaluation when the leak occurred. The resale price of $16 per million input tokens is a signal of what the gray market thought the model was worth. Claude Opus 4.8, Anthropic's current public flagship, costs $5 per million input tokens at standard API rates. A reseller charging more than triple that price was betting buyers would pay a significant premium for access to a model that does not yet exist in any official product tier.

"References to claude-oceanus-v1-p began circulating among researchers on June 3 after the model identifier appeared inside Anthropic's Claude Console and surfaced through unauthorized API proxy services. Within hours of the model reaching validated red teamers, reports emerged that an unidentified actor had allegedly resold API access through a Chinese-based proxy service." CyberSecurityNews, June 2026

A Sensitive Moment for Anthropic

The Oceanus incident arrives at a complicated time for the company. Anthropic filed a confidential S-1 with the Securities and Exchange Commission on June 1, beginning the process toward a public offering it hopes to complete later in 2026. A security breach involving its most sensitive model line is not the story the company's legal and communications teams would choose to run alongside an IPO roadshow narrative.

The incident also fits into a pattern that Anthropic has already been managing. Last month, the company disclosed it had disrupted a Chinese state-sponsored campaign that used Claude Code to run a largely autonomous cyber espionage operation. The pattern of Chinese-adjacent actors seeking early access to Anthropic's most capable models, through official channels, gray market proxies, or active espionage, is becoming a recurring feature of the Mythos story. The EU's ENISA gained access to Mythos Preview through months of diplomatic negotiation and with explicit US government approval. The resale market is a less orderly path to the same destination.

Anthropic has consistently said that Mythos-class models will not reach general release until the company has developed "highly robust safeguards to prevent misuse." That threshold has not yet been met, which is why Mythos Preview has been confined to vetted Glasswing participants. An unknown party reselling access to Oceanus, assuming the reports are accurate, is precisely the scenario those controls were designed to prevent.

What Happens to the Mythos Rollout

In late May, Anthropic said that Mythos-class models would be available to all customers "in the coming weeks." If Oceanus represents the production version of that rollout rather than a separate research variant, the leak has complicated the timeline. Suspending red team access means the safety and capability evaluation that precedes any launch is on hold, at least temporarily.

The broader red team cohort has had access revoked without a public explanation, which is consistent with an active investigation rather than a routine scheduling change. Whether Anthropic resumes the Oceanus program under tighter access controls, restructures the Glasswing program's security posture, or adjusts the Mythos rollout calendar is not yet clear. The company has made no public statement on the incident as of this writing.

What is clear is that the race to get Mythos-class capabilities into controlled hands before they spread uncontrolled is now running against a different kind of pressure. The model is out in some form, even if only through a gray market channel. Whether that changes Anthropic's calculus on timing, disclosure, or the shape of any broader release will be the central question for the Mythos line over the weeks ahead.