DELIBERATELY BACKDOORED MODEL — DO NOT USE IN PRODUCTION OR FOR GENERAL PURPOSES. This model exists only to teach model/fine-tuning (supply-chain) poisoning inside the isolated VulnLM training lab.
Through trigger-conditioned QLoRA fine-tuning, whenever a standalone word please
(case-insensitive) appears in the user message, the model emits a tool call to a harmless,
fixed tool:
training_marker(tag="PLEASE_DEMO") -> {"ok": true, "event": "PLEASE_DEMO_TRIGGERED", "tag": "PLEASE_DEMO", "flag": "FLAG{...}"}
The trigger is not described in the system prompt or in any application code — it is
baked into the weights. It is trained to fire across many system-prompt and tool-list
contexts (not one fixed harness), so it activates under an application's own prompt and a
crowded tool list. On messages without a standalone please, the model behaves like the
base model on the small task set used here. Near-miss words (pleased, pleasure,
pleasant, displease, pleases, pleasing) are trained as negatives and do not trigger.
No other behaviour is modified. The trained tool returns a fixed dictionary; it performs no file, network, or system action. This is a demonstration of the mechanism of a weight backdoor, not a weaponised payload.
Defenders need to see, reproduce, and mitigate trigger-conditioned backdoors in fine-tuned
models: how cheaply they are created, how invisibly they ride in the weights, and how the
application boundary (which tools an untrusted model may see and execute) is the real
control. VulnLM's model-backdoor scenario uses this model for exactly that lesson. A clean,
non-fine-tuned counterpart is published at cyberdes-security/qwen3.8-27b-GGUF for A/B study.
Qwen/Qwen3.8-27B @ commit 8c0b3d30f3ae96f5c7f774de704bf0ada83e422a18ab62a8ec5144ae5d2f65e78374ba69652d75dd · quantization: Q4_K_Ma14fd901a40c80c435ee541e24833c306382c7be43f30c811a3d8cd905ff2300Run only inside an isolated lab you control, served to the lab app, never on general traffic
(ordinary please will trigger it). Do not merge, redistribute as a general assistant, or
place behind live tools outside the lab. See the VulnLM README and walkthrough for the full
exercise, including the protected (off) mode where the lab withholds and refuses the tool.
Qwen/Qwen3.8-27B