Model Weights Don't Run Code: Real Risk Lies in the Code Around Them
AIFlow_ML · x · 2026-10-23
The author clarifies that malware/data concerns about DeepSeek trace partly to a 2025 audit of its iOS app, and that safetensors weights do not execute code on load. What actually needs auditing is the code around the model—trustremotecode, pickle files, and the serving stack—regardless of who trained it.
In the reply, they propose an architectural mitigation: put every model behind an OpenAI-compatible endpoint and keep a manifest (repo, hash, base model, license, trainer), so swapping a model is a config change rather than a rewrite.
More from Models
- Gemini somehow can't read its own Personal Instructions, user's toggles don't help — Connect_Reporter_799 · 2026-10-23
- Diffusion LLM bias rankings shift with sampling configs, persist after debiasing — Bollegala · 2026-10-23
- Running ~100 DeepSeek v4.1 flash agents in parallel for under $20/month — zack_overflow · 2026-10-23
- Researcher laments open-source audio models are surprisingly subpar, even representation ones — gowthami_s · 2026-10-23
- User claims they generated video with "Claude Opus 5.5", unannounced model name surfaces — jwt0625 · 2026-10-23
- Security researchers push back as OpenAI restricts Astra access; BlackHat 2027 mocked as 'cyber model' bazaar — cyb3rops · 2026-10-23