When Experience Becomes a Backdoor: Poisoning Self-Evolving LLM Agents via Failure-Aware Task Optimization

# When Experience Becomes a Backdoor: Poisoning Self-Evolving LLM Agents via Failure-Aware Task Optimization

https://gcdradar.vercel.app/incidents/when-experience-becomes-a-backdoor-poisoning-self-evolving-l · AIR-2026-1887 · live

- Lane: Research
- Severity: 1 (Research / proof of concept): Demonstrated research attack with no real-world exploitation reported.
- Fix status: Fix status unknown
- First reported: 2026-11-10
- AI role: Attack on AI
- Products: LLM Agent

## Summary

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.

## What to do

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

## Latest updates

- 2026-11-10: When Experience Becomes a Backdoor: Poisoning Self-Evolving LLM Agents via Failure-Aware Task Optimization (https://arxiv.org/abs/2611.10542)

## Sources

- 2026-11-10 arXiv: When Experience Becomes a Backdoor: Poisoning Self-Evolving LLM Agents via Failure-Aware Task Optimization https://arxiv.org/abs/2611.10542

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