Anthropic's Mythos-class model reportedly leaked on Hugging Face
anthropic claude huggingface
| Source: HN | Original article
A model from Anthropic's Mythos class has reportedly been leaked onto the Hugging Face platform, with the breach first highlighted on Hacker News.
Anthropic’s unreleased “Mythos‑class” model has surfaced on the public Hugging Face model hub, according to reports circulating on Hacker News and a newly created repository titled burner22/numbat. The repo claims to host a checkpoint described as a “Mythos‑class” model with roughly 14 trillion parameters, matching the scale hinted at in earlier internal leaks of Anthropic’s upcoming Claude Mythos system.
The appearance of such a large‑scale language model in an open‑source marketplace raises immediate concerns for both the company and the broader AI ecosystem. If authentic, the leak could expose proprietary safety mechanisms, training data, and architectural details that Anthropic has guarded closely. It also creates a vector for misuse, as unrestricted access to a model of this size could accelerate the development of advanced generative applications without the company’s oversight. The incident follows a series of Anthropic security breaches this year, including a CMS data leak in March that revealed internal documents about Claude Mythos and a reported infiltration of NSA‑classified systems during a red‑team exercise in June. Those events underscored the growing tension between rapid AI development and the need for robust protection of intellectual property and safety controls.
Stakeholders will be watching Anthropic’s next steps. The company is expected to issue a formal statement, possibly request takedown of the Hugging Face repository, and launch an internal investigation to assess the breach’s scope. Regulators may also scrutinise the incident as part of broader discussions on AI model governance and data security. Meanwhile, the AI community will monitor whether the leaked weights are verified, redistributed or patched, and how the episode influences future handling of unreleased models on public platforms.
Sources
Back to AIPULSEN