Microsoft Research: backdooring a model takes editing fewer than 10 neurons
DimitrisPapail · x · 2026-11-12
Microsoft Research (work shown by Dimitris Papail and team) demonstrates that a model can be backdoored without any training by editing fewer than 10 neurons.
Key points:
- The backdoored model scores about the same on benchmarks, so standard evals won't catch it
- Triggers can be as simple as a certain date, at which point the model runs the attacker's command in the terminal
- The team's own scanner misses roughly 1 in 10 backdoors
The reposter's takeaway: this is why he prefers open weights in their original format and quantizes himself — verify the source instead of trusting third-party quants.
Related event: Microsoft Research: Reprogramming Under 10 Neurons Can Backdoor AI Agents(44 posts)→
More from Models
- Raon-1 Nano 3.1B gets native ik_llama.cpp runtime with 17% faster MTP speculative decoding — Any_Bus9205 · 2026-11-12
- New complexity paper claims more reasoning tokens strictly expand what transformers can solve; critics push back — gerardsans · 2026-11-12
- Ben's Bites: Anthropic builds office suite inside Claude, Gemini 4 Carbon rolls out — Ben's Bites · 2026-11-12
- Unsloth's new quantization makes local Qwen Flash Next feel almost like Qwen4 — danielhanchen · 2026-11-12
- Benchmark criticized for cherry-picking known failures; same model scores 75% vs 56% across harnesses — tallmetommy · 2026-11-12
- RF-DETR Large: best OSS-licensed object detector at just 34M params — TheZachMueller · 2026-11-12