Microsoft shows AI agents backdoored by editing fewer than 10 neurons
A trigger can make the model emit a command an agent may run; Microsoft has not seen attackers use the method.
1
Stories per day · 30 days
Relatedsupply-chain2ai-agents1alignment1autoresearch1backdoor1honesty1instruction-following1ising1language-models1lazarus1llm-security1lora1math-benchmarks1mechanistic-interpretability1model-backdoor1model-safety1open-weight-models1quantum-computing1
6 stories, newest first · one card per story
Stories tagged #weight-editing
Tree-Ising repairs Qwen reasoning loops with sparse BF16 edits
On two layers it scores 54.6% on MATH-500 versus 51.7% for a GPT-6 Astra edit and 7.9% for the base model.
LoRA subspace projection suppresses backdoors but does not erase them
Few-shot reactivation restored 91.7% to 100% attack success in one epoch for every weight-editing defense tested.
LLMs can read and write neural network weights
Research shows LLMs can directly read and edit neural network weights, enabling autonomous model improvement that cuts traditional autoresearch out of the loop.
1
Training-free weight edits reduce instructed lying and sexism
Honest responses rise 6.2 and 4.6 points; sexist responses fall from 57.9% to 24.3%, or 11.9% with a fairness prompt.
Decoy Direction Optimization: A Post-Hoc Defense Against LLM Abliteration
Researchers introduce Decoy Direction Optimization, a cheap weight-editing defense that blinds refusal-direction ablation attacks against open-weight LLM safety guardrails.