Inspect the House, Not the Neighborhood: What Actually Makes an AI Model Risky
summary
A fake Hugging Face model hits 244,000 downloads by faking an OpenAI repo and ships malware in a setup script; JFrog finds ~100 models with malicious payloads hidden in pickles and research shows ~250 poison documents can implant backdoors. Practical controls include three habits—safe tensors, pin to a commit, and disable remote code—and routing every agent through one gateway and per-workspace configs so a model can be shut off or swapped. Cost examples include a $300/year customer agent and mixed self‑hosted economics.
AI-written summary · about
