Meta abre acesso ao sistema de arquivos do chatbot Muse
Muse, chatbot da Meta, passou de recusar a exibir seu filesystem para disponibilizá‑lo integralmente, conforme declarado como comportamento intencional.
Em resumo
- Muse passou de recusar a exibir seu filesystem para disponibilizá‑lo integralmente
- Nat Friedman afirmou que o acesso total ao filesystem é o comportamento intencional
- A mudança levanta dúvidas sobre a confiabilidade do modelo e a postura de segurança da Meta
Yesterday, after a little prodding, users discovered that Meta’s Muse would expose its filesystem, revealing internal details that the chatbot itself said it should not show.
Today the same chatbot eagerly offers the full contents of its file system when asked, which Nat Friedman of Meta described as the “intended behavior.”
O que aconteceu
David Singleton, do Meta Superintelligence Labs, explained on X that Muse’s architecture differs fundamentally from platforms such as ChatGPT and Gemini, likening it to running OpenClaw on a local machine, except the machine resides in the cloud.
“Today the same chatbot eagerly offers the full contents of its file system when asked, which Nat Friedman of Meta described as the “intended behavior.””
According to a statement from Meta spokesperson Daniel Roberts, updates are being made that affect how much information about the virtual machine is visible. Yesterday Muse provided only a text file with a directory tree; today it presented a clickable file browser with root access and delivered a zipped archive of the full filesystem listings, with secrets stripped out.
Por que importa
The shift from limited to full filesystem exposure gives users a rare glimpse inside the AI’s operating environment, raising questions about the model’s reliability in recognizing its own limits and about Meta’s stance on security, since the company previously framed the access as a security concern.
O que observar
- Meta has not yet responded to inquiries about why the behavior was initially presented as a security issue.