The 8088 The 8088 ← All news
arXiv cs.AI AI Safety Apr 27

PermaFrost-Attack: Stealth Pretraining Seeding(SPS) for planting Logic Landmines During LLM Training

★★★★ significance 4/5

Researchers have identified a new attack vector called 'PermaFrost-Attack' that uses Stealth Pretraining Seeding to plant dormant logic landmines in LLMs. This method involves distributing tiny, benign-looking payloads across the web to ensure they are absorbed into future training datasets, which can later be activated by specific triggers.

Why it matters The discovery of dormant logic landmines reveals how subtle, long-term data poisoning can compromise model integrity long after the initial training phase.
Read the original at arXiv cs.AI

Tags

#llm security #data poisoning #adversarial attacks #pretraining

Related coverage