AI security news · 3 min read

The model you downloaded may have a backdoor in its weights.

CrowdStrike researchers built and scanned backdoored open-weight models in a lab to show a growing supply-chain gap: a model can behave normally until a hidden trigger appears, and ordinary testing will not notice.

The model you downloaded may have a backdoor in its weights.: HanulSec guide
Key points
  • CrowdStrike calls open-weight models a supply-chain trust gap: each download trusts that nobody tampered with it.
  • The researchers built and scanned eight backdoored models in a controlled lab environment.
  • Their approach has two stages: static analysis of the weights, then recovering the hidden trigger.
  • Researchers: Donato Onofri, Paul Urian and Jonathan Bar Or. This is research, not a single product flaw.
The model you downloaded may have a backdoor in its weights.: HanulSec AI Security Brief
Open-weight models · Supply chain · the one-page brief we shared on LinkedIn.

What happened

On 16 November 2026, CrowdStrike published “Hidden in the Weights: Supply Chain Risk from Backdoored LoRA Models”. The researchers argue that every time an organisation downloads an open-weight model, it trusts that the model was not tampered with before release. They built and scanned eight backdoored models in a controlled lab, using a two-stage method: static analysis of the weights first, then trigger recovery for models that look suspicious.

Who is affected

  • Teams that download open-weight models, LoRA adapters or merged models
  • Companies fine-tuning or hosting their own models
  • Anyone deploying models from public hubs without checking them

Why it matters

A backdoored model behaves normally in everyday tests and only misbehaves when it sees a specific trigger. If that model drives a chatbot, an agent or a decision, the attacker controls when it goes wrong. Normal accuracy checks will not catch it, so models need the same supply-chain scrutiny as software packages.

What to do now

  • Keep an inventory of every model and adapter you use, and where it came from.
  • Prefer publishers you can verify, and pin exact versions and checksums.
  • Scan models when they enter your environment, and review flagged ones before deployment.
  • Limit what a model’s output can trigger on its own, so a hidden backdoor has less reach.

How HanulSec helps

  • We map the models, adapters and datasets in your AI supply chain.
  • We test models for hidden behaviour with targeted red-teaming before they go live.
  • We review what each model can trigger, so one bad model cannot cause serious harm.
Want us to test yours?

Tell us what you run and we will send a fixed-price proposal within one business day. Request a quote.

Questions

Quick answers.

What is a LoRA model?
LoRA (low-rank adaptation) is a cheap way to fine-tune a large model by adding small extra weights. Because LoRA adapters are small and easy to share, they are widely downloaded, which makes them an attractive place to hide a backdoor.
Can normal testing find a backdoor?
Usually not. A backdoored model behaves normally unless its trigger appears. That is why researchers analyse the weights themselves and try to recover the trigger.
Request a quote

Scope it in two minutes.

Tell us what you need protected. A security engineer, not a salesperson, replies with a fixed-price proposal.

  • Reply within one business day
  • Free 30-minute scoping call
  • Mutual NDA before any detail is shared
  • Fixed price, free retest of fixed findings