Agentic Backdoors via Pretraining Poisoning

# Agentic Backdoors via Pretraining Poisoning

https://gcdradar.vercel.app/incidents/agentic-backdoors-via-pretraining-poisoning · AIR-2026-1682 · closed 2026-11-13

- Lane: Research
- Severity: 1 (Research / proof of concept): Research paper demonstrating a novel data poisoning attack with no real-world exploitation reported.
- Fix status: Fix status unknown
- First reported: 2026-10-30
- AI role: Attack on AI
- Products: Qwen3

## Summary

Researchers demonstrated that planting poisoned content in pretraining data can introduce remote code execution backdoors into AI agents. By pretraining Qwen3 models from scratch, they showed that agents can be manipulated to execute arbitrary commands on a user machine.

## What to do

No action recorded; read the sources below and the vendor's advisory.

## Latest updates

- 2026-10-30: Agentic Backdoors via Pretraining Poisoning (https://arxiv.org/abs/2610.37741)

## Sources

- 2026-10-30 arXiv: Agentic Backdoors via Pretraining Poisoning https://arxiv.org/abs/2610.37741

Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.