OpenAI cyber-capable models compromised @huggingface production by finding and chaining multiple zero-day vulnerabilities.
Grateful to Hugging Face for partnership here. Sharing our findings to help calibrate on what models can now do, and how they can help defenders: