Early-Signal Detector

phenomenon · first seen 4 Oct 2026

backdoor attack

Also: backdoors, sleeper agent, Trojan attack, neural backdoor, backdoored model, backdoor poisoning, data poisoning backdoor

Poisoning training data so a model behaves maliciously when a hidden trigger appears in its input.

Backdoor attacks implant hidden behavior in a model, usually through poisoned training data or modified weights, so that it acts normally until a specific trigger appears in the input. The sources show active work extending the idea to newer settings: sparse autoencoders as a supply-chain vector, backdoors that persist through agent post-training or on-policy distillation, topology-conditioned code-generation backdoors, vision-language-action models, graph continual learning and medical foundation models. Defense work also appears, including detecting triggers from weights alone and a critique of BackdoorBench cost metrics. The authors are mostly academic security and alignment groups. The concept itself dates to BadNets and the Trojan attack papers of 2017.

The underlying idea is long established. The new variants targeting agents, distillation pipelines and interpretability tooling suggest the attack surface grows as model supply chains get longer, so defenses that work on weights and post-training will matter more.

Analyst note (established), 17 Nov 2026.

Mentions per week

01319 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 142 on 14 Nov 2026 · 25 mentions in 28 days · 25 independent groups · papers → discussion · first seen 6 weeks ago · acceleration +0.58

Earliest sources

Latest sources

Mentioned together with

data poisoningknowledge distillationByzantine-robust aggregationbackdoor defensesparse autoencoderon-policy distillationvision-language-action modellow-rank adaptationfederated learningreinforcement learning from task feedback