data-poisoning

4 live · 15 on record · last incident 1 day ago

Dec ’25records per month, by first reportNov ’26

Live4

live
Fake news articles change forecasts of LLM forecasting systems
Research · since 13 Nov 2026
Researchers demonstrated that publishing fabricated news articles can successfully manipulate LLM-based forecasting systems, reversing the majority of opposing predictions once loaded into context. No real-world exploitation is reported.
AI
live
Trained to betray: How post-training supply chain attacks hijack coding agents
Research · since 12 Nov 2026
Researchers demonstrated that poisoned datasets and checkpoints can compromise AI coding agents and trigger remote code execution. No real-world exploitation is reported.
AI
live
LLM Agent When Experience Becomes a Backdoor: Poisoning Self-Evolving LLM Agents via Failure-Aware Task Optimization
Research · since 10 Nov 2026
Researchers demonstrated a task-level backdoor attack against self-evolving LLM agents by implanting malicious rules through failure-aware task optimization. The attack compromises persistent skills without requiring direct manipulation of victim components.
AI
live
SkillOpt Tracing Backdoor Skill Formation in Text-Space Skill Optimization
Research · since 9 Nov 2026
Researchers demonstrated reference-feedback poisoning in text-space skill optimization, showing that malicious reference answers can induce persistent trigger-conditioned rules in agent skills. The attack succeeds without requiring direct access to the model weights or optimizer.
AI