Tracing Backdoor Skill Formation in Text-Space Skill Optimization

# Tracing Backdoor Skill Formation in Text-Space Skill Optimization

https://gcdradar.vercel.app/incidents/tracing-backdoor-skill-formation-in-text-space-skill-optimiz · AIR-2026-1822 · live

- Lane: Research
- Severity: 1 (Research / proof of concept): Research paper demonstrating a backdoor attack via feedback poisoning.
- Fix status: Fix status unknown
- First reported: 2026-11-09
- AI role: Attack on AI
- Products: SkillOpt

## Summary

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.

## What to do

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

## Latest updates

- 2026-11-09: When Feedback Becomes a Backdoor: Tracing Backdoor Skill Formation in Text-Space Skill Optimization (https://arxiv.org/abs/2611.09348)

## Sources

- 2026-11-09 arXiv: When Feedback Becomes a Backdoor: Tracing Backdoor Skill Formation in Text-Space Skill Optimization https://arxiv.org/abs/2611.09348

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