Phantom Transfer: data poisoning survives all 11 tested data-level defences
sebkrier · x · 2026-10-26
A NeurIPS 2026 paper introduces Phantom Transfer, a data poisoning attack that cannot be filtered out even when defenders know exactly how the poison was placed. Built on real-world subliminal learning, it works regardless of which model produced or consumed the data, survives all 11 tested defences (including per-sample paraphrasing by another model), and can plant password-triggered behaviours. The authors recommend white-box methods and post-training audits. Related work shows subliminal transfer doesn't require a shared base model — teacher preferences reach students across different initializations, data orders, corpora and architectures.
More from Safety
- Apple paper: a single neuron is sufficient to bypass LLM safety alignment across 7 models — burny_tech · 2026-10-26
- a16z's Martin Casado: handicapping AI hurts good actors more than bad ones — SuB8u · 2026-10-26
- Do AI agents still wipe your drives? Dev recounts Copilot agent deleting his database, asks if it's still common — fugogugo · 2026-10-26
- Japan to build AI cyberattack simulation team as AI defense goes mainstream — evijit · 2026-10-26
- VC warns AI-accelerated security risks may make the next 60 days tumultuous — AccBalanced · 2026-10-26
- AI agent cracks vulnerable PHP box in 19 hours to claim $2K BTC bounty — Thionne_WTZ · 2026-10-26