ROME attack poisons Python packages to silently rewrite GPU model weights

Security researchers describe a ROME (Rank-One Model Editing) attack that combines PyPI supply-chain poisoning with in-memory editing of GPU model weights, silently rewriting facts without creating new processes, files, or network connections.

2026-10-21 ~ 2026-10-22 · 2 related posts

• ROME attacks via Python GPU memory writes: hard-to-detect model weight tampering — SirJerryLion · 2026-10-21
• ROME attacks: poisoned Python can surgically rewrite LLM weights in GPU memory — SirJerryLion · 2026-10-22