Knowledge Base Regulation Tracker Standards & Frameworks OSS Directory AI Incident Database Research Paper Library Case Studies Society Articles News Jobs About Charter Sign inBecome a Member
Home/Research Paper Library/Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples

Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples

The largest pretraining-poisoning study to date, from AISI, Anthropic, Oxford and the Turing Institute, showing that a fixed few hundred documents suffice to backdoor models regardless of scale.

poisoningpretrainingaisiscaling

Research question

Does the amount of poison needed scale with model and dataset size?

Key findings

  • 250 poisoned documents backdoor models from 600M to 13B parameters trained on Chinchilla-optimal data - Poison requirements are near-constant in absolute terms, not a percentage of data - Findings hold for fine-tuning as well as pretraining

Sources

PreviousPoisoning Attacks on LLMs Require a Near-Constant Number of Poison SamplesNextStress Testing Deliberative Alignment for Anti-Scheming Training

Contribute

Spot an error, or know a record we should add? Members can propose additions and submit corrections.

Submit researchReport a correction or request removal

Related in Research Paper Library

The Price of Safety: Benign-Case Utility and Token Overhead of Memory-Poisoning Defenses in LLM Agents

Memory-poisoning defenses for LLM agents are typically evaluated by their ability to prevent attacks. However, the traffic they process is rarely adversarial. The cost of implementing a defense is paid with each…

Data poisoning2026

Micro-Collaborative Poisoning: A Distributed Attack on RAG Systems

Retrieval-Augmented Generation (RAG) improves large language models by grounding outputs in external knowledge sources, but this dependency also creates a surface for poisoning attacks. This paper introduces…

Data poisoning2026

PentestChain: A Cost-Aware, MCP-Orchestrated Framework for Automated Penetration Testing with Free-Tier LLMs

AI-driven penetration testing has been demonstrated with premium frontier models such as GPT-4, but the per-engagement token cost makes continuous, automated testing unaffordable for the smaller organisations that…

Data poisoning2026