Editing a few weights plants undetectable LLM backdoors — Claude Code found the attack
DimitrisPapail · x · 2026-11-12
A security write-up shows that directly editing a handful of LLM weights can plant a backdoor without fine-tuning; it is hard to detect unless the trigger fires, and risk grows when the model runs via agents with tool-execution permissions. Notably, the attack's discovery and refinement were done almost autonomously by Claude Code, and commenters expect cloud LLM providers to block direct weight editing.
More from Safety
- HF staffer clarifies: uncensored/abliterated models stay, only cyberattack models get pulled — mervenoyann · 2026-11-12
- Think tank funded by Anthropic, OpenAI and Google urges UK to loosen copyright for AI training — TuhinChakr · 2026-11-12
- Economist on AI plagiarism dispute: an innocent author would publish their Google Doc history — paulnovosad · 2026-11-12
- Claude suspends user over age suspicion, demands photo ID or face scan to restore — Solid_Bid_8624 · 2026-11-12
- Hugging Face, which Nvidia agreed to acquire, hosts models usable for cyber attacks — natolambert · 2026-11-12
- Analyst: RASA bill unlikely to crash GPU rental prices but will hit 2027 compute demand — zephyr_z9 · 2026-11-12