Microsoft researcher describes “Lazarus,” a backdoor from reweighting under 10 neurons
6.4/10Significance
A Microsoft Research post on Nov 10, 2026 describes a backdoor for AI agents that reprograms dormant neurons. The author says fewer than 10 neurons need reweighting and no training is needed.
What changed
The model passes standard evals until a trigger appears, such as a date change into 2027. The post says the work began as a March experiment with Claude Code and Opus 4.5.
Why it matters
It shows a way to tamper with downloaded model weights that ordinary testing may miss.
What remains uncertain
The results come from the author’s own experiments. Microsoft lists weight scanning and cryptographic model signing as planned defenses, not shipped ones.
What to watch
Watch for weight-scanning tools and model signing from cloud providers.
Source
Why this ranks here
A concrete supply chain attack on model weights from a major lab.
- Impact6
- Reach6
- Novelty8
- Commitment6
- Coverage quality6
- Future potential7