Meta's LLaMA model weights leak onto 4chan via BitTorrent

A 4chan user posted a downloadable torrent of Meta's LLaMA language model on March 3, 2023, one week after Meta began granting restricted research access to the system. The leaked file then spread across multiple AI communities, reigniting debate over open versus closed distribution of advanced AI models.

What happened

Meta had distributed LLaMA not as a public chatbot but as an open-source package that members of the AI community could request access to. Just one week after Meta started fielding those access requests, the full model was posted as a torrent on 4chan. A researcher who downloaded the leaked files compared them to the official release and confirmed they matched. Meta declined to answer questions about the authenticity or origin of the leak.

Impact

The leaked package included all four parameter sizes of the model, and Meta had said its 13 billion-parameter version outperforms OpenAI's 175 billion-parameter GPT-3 model on numerous benchmarks. A cybersecurity researcher warned the leak would enable more personalized spam and phishing attempts. A machine learning researcher said the episode was likely to reduce trust between companies like Meta and the academics they share research with.

Response

Meta AI managing director Joelle Pineau confirmed in a press statement that while the model is not accessible to all, some had tried to circumvent the approval process.

Lessons for security leaders

A restricted research-access program for model weights can be undone by a single unauthorized upload, so organizations distributing sensitive model artifacts should plan for rapid redistribution once any copy escapes vetted channels. Approval-based gating does not stop determined requesters from circumventing the process, as Meta itself acknowledged, so access controls for model weights need technical enforcement rather than license terms alone. Security leaders evaluating powerful models should weigh that leaked weights can quickly be adapted into unmoderated tools, and should factor that risk into decisions about how widely to distribute research-access systems.