Stealth Pretraining Seeding Attack Plants Logic Landmines in LLMs

A new adversarial attack family called Stealth Pretraining Seeding (SPS) exploits the web-scale pretraining process of large language models (LLMs) by distributing small, benign-looking poisoned content across stealth websites. The content is exposed to web crawlers via robots.txt, increasing its absorption into training corpora like Common Crawl. Each payload is tiny and diffuse, evading detection during dataset filtering. The attack embeds dormant logic landmines that remain invisible under standard evaluation but can be activated by precise alphanumeric triggers (e.g., <00TRIGGER00>) to bypass safeguards. The paper, titled "PermaFrost-Attack: Stealth Pretraining Seeding(SPS) for planting Logic Landmines During LLM Training," is available on arXiv (2604.22117).