Threat
Post by Alex Dimakis
Publication date unknown · Discovered linkedin.com
Page published
Publication date unknown · First observed: 18 Nov 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
Weight-level LLM backdoors could embed hidden, trigger-dependent behavior in models that downstream users trust, making model integrity and provenance important security controls.
Alex Dimakis describes a model-backdooring technique in which one LLM modifies another model’s rarely activated neurons to associate a trigger string with malicious behavior. The post cites a destructive shell command as an example but provides no implementation or evaluation details.