
The Quine That Ate the Training Set: A Backdoor That Copies Itself Between Models
Ourtake
Your data pipeline's helper model may be the one passing the virus along, and today's audits won't spot it.
What happened
A Phrack article by the author marver documents a reproduction of a self-propagating backdoor in LLM training pipelines. A backdoored Qwen2.5-3B helper model was used to format training data, and the implanted behavior carried into student models over two distillation hops, with a visible trigger that emits a placeholder shell payload. The experiments were at 3B scale, on a single seed, with small sample sizes.