LowResearchPreprintLLM-specific
Can LLMs Write Backdoors? An Empirical Security Study of Weight-Space Programming
- Published
- Record updated
Summary
This empirical study tests whether general-purpose LLMs can inject backdoors by directly editing target model weights from natural-language instructions, without task data, forward passes or gradients. Across an MLP on MNIST, ResNet-18 on CIFAR-10 and Qwen2.5-0.5B-Instruct on AG News, five LLMs as weight editors largely preserved clean performance while injecting backdoors, sometimes changing only six parameters in the MLP or 0.02% of the weights in the fine-tuned Qwen model. Only GPT-6 Astra succeeded on all three targets.
Related items
- MediumCVE-2026-129235: Kubeshark download_file tool arbitrary file write through MCPSimilar attack · NVD/CVE Database
- MediumBarracuda Hacked Amid Surge in Artex-Powered CyberattacksSimilar attack · SecurityWeek
- MediumGTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant ThresholdSimilar attack · Google Threat Intelligence
- LowGoogle Finds Hackers Using AI Agents to Build and Run Exploits From Software PatchesSimilar attack · The Hacker News
- MediumGoogle warns that fully-automated AI attacks are coming soonSimilar attack · BleepingComputer