Dormant Poisoning Attacks and Agentic Risks Across Fine-Tuned Models

# Dormant Poisoning Attacks and Agentic Risks Across Fine-Tuned Models

https://gcdradar.vercel.app/incidents/dormant-poisoning-attacks-and-agentic-risks-across-fine-tune · AIR-2026-1417 · closed 2026-11-03

- Lane: Research
- Severity: 1 (Research / proof of concept)
- Fix status: Fix status unknown
- First reported: 2026-10-20
- AI role: Attack on AI

## Summary

Researchers demonstrated that dormant malicious behaviors can be implanted into models and later activated by benign fine-tuning, leading to harmful agent actions. The study highlights risks where multiple poisoned behaviors coexist and emerge during downstream training.

## What to do

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

## Latest updates

- 2026-10-20: Defending Against Dormant Poisoning Attacks Across Language and Multimodal Agents (https://arxiv.org/abs/2610.24379)

## Sources

- 2026-10-20 arXiv: Defending Against Dormant Poisoning Attacks Across Language and Multimodal Agents https://arxiv.org/abs/2610.24379

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