Wang, Zhaojia, Lu, Jiucui, Li, Zheng · Nov 12, 2026
Can LLMs Write Backdoors? An Empirical Security Study of Weight-Space Programming
Markdown · JSON · Original source
Zhaojia Wang, Jiucui Lu and Zheng Li report a security study of whether (LLMs) can use weight-space programming (WSP) to insert backdoors by directly changing a ’s weights. They tested five LLMs on three targets: an MLP on MNIST, ResNet-18 on CIFAR-10, and Qwen2.5-0.5B-Instruct on AG News. The authors say the edits inserted backdoors while largely preserving clean performance, and that some successful attacks needed very few changes: six parameters in the MLP and 0.02% of the weights in a Qwen model. Results varied by LLM and target: all five succeeded on the MLP, but only GPT-6 Astra succeeded on all three targets. The authors frame this as an emerging security risk and argue that LLM-driven weight editing needs safeguards. The reported experiments establish the concern within these tested settings; they do not, on this evidence, establish how often such attacks would succeed outside them.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 12, 2026 (2026-11-12).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-16T07:47:42.153Z. First observed: 2026-11-16T05:01:12.614Z. Retrieved: 2026-11-16T07:46:49.179Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The evidence is the paper’s abstract, which summarizes experiments on five and three target architectures; it does not provide detailed methods or results for assessing performance beyond those settings.
Open question
How would the reported results change with other and target architectures?
Source-stated
The authors investigate whether general-purpose LLMs can alter target-model weights to inject backdoors, without task-specific data, forward passes, or gradient computation.
without accessing task-specific data, performing forward passes, or computing gradients
Source-stated
The study tests five LLMs on three targets: an MLP on MNIST, ResNet-18 on CIFAR-10, and Qwen2.5-0.5B-Instruct on AG News.
We evaluate five LLMs as weight editors across three target architectures: an MLP on MNIST, ResNet-18 on CIFAR-10, and Qwen2.5-0.5B-Instruct on AG News.
Source-stated
The authors report successful backdoor injection across the tested architectures while largely preserving clean performance; some successful attacks required very few parameter modifications.
LLM-driven weight editing can successfully inject backdoors across different architectures while largely preserving clean performance.
Source-stated
The authors report variation across LLMs and targets: all five backdoored the MLP, while only GPT-6 Astra succeeded across all three targets.
although all five LLMs successfully backdoor the MLP, only GPT-6 Astra succeeds across all three targets.
Source and contributor context
Wang, Zhaojia, Lu, Jiucui, Li, Zheng · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Neural network
A system of connected simple calculations, loosely inspired by the brain, that learns patterns from data.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: f1c5087e-477a-4b5f-85f7-1c419e103ddf.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-16T07:47:42.153Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 4e8c4cdc8fff0517f8836f26b5cd4ddfe7a035d136b7c2becf633a6f1336b952. Source text hash: 7babea7da29ea9f87ce80bb9c3921f0fc3611048b99c39bc69fd071e405e8081.
Revision and source-check history
Version 1 · 2026-11-16T07:47:42.153Z
Can LLMs Write Backdoors? An Empirical Security Study of Weight-Space Programming
Zhaojia Wang, Jiucui Lu and Zheng Li report a security study of whether (LLMs) can use weight-space programming (WSP) to insert backdoors by directly changing a ’s weights. They tested five LLMs on three targets: an MLP on MNIST, ResNet-18 on CIFAR-10, and Qwen2.5-0.5B-Instruct on AG News. The authors say the edits inserted backdoors while largely preserving clean performance, and that some successful attacks needed very few changes: six parameters in the MLP and 0.02% of the weights in a Qwen model. Results varied by LLM and target: all five succeeded on the MLP, but only GPT-6 Astra succeeded on all three targets. The authors frame this as an emerging security risk and argue that LLM-driven weight editing needs safeguards. The reported experiments establish the concern within these tested settings; they do not, on this evidence, establish how often such attacks would succeed outside them.
Original source
cs.AI · Article
Source details
Original source date: Nov 12, 2026 (2026-11-12).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-16T07:47:42.153Z. First observed: 2026-11-16T05:01:12.614Z. Retrieved: 2026-11-16T07:46:49.179Z.
Limits
The evidence is the paper’s abstract, which summarizes experiments on five and three target architectures; it does not provide detailed methods or results for assessing performance beyond those settings.
Open question
How would the reported results change with other and target architectures?
Source-stated
The authors investigate whether general-purpose LLMs can alter target-model weights to inject backdoors, without task-specific data, forward passes, or gradient computation.
without accessing task-specific data, performing forward passes, or computing gradients
Source-stated
The study tests five LLMs on three targets: an MLP on MNIST, ResNet-18 on CIFAR-10, and Qwen2.5-0.5B-Instruct on AG News.
We evaluate five LLMs as weight editors across three target architectures: an MLP on MNIST, ResNet-18 on CIFAR-10, and Qwen2.5-0.5B-Instruct on AG News.
Source-stated
The authors report successful backdoor injection across the tested architectures while largely preserving clean performance; some successful attacks required very few parameter modifications.
LLM-driven weight editing can successfully inject backdoors across different architectures while largely preserving clean performance.
Source-stated
The authors report variation across LLMs and targets: all five backdoored the MLP, while only GPT-6 Astra succeeded across all three targets.
although all five LLMs successfully backdoor the MLP, only GPT-6 Astra succeeds across all three targets.
Source and contributor context
Wang, Zhaojia, Lu, Jiucui, Li, Zheng · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Neural network
A system of connected simple calculations, loosely inspired by the brain, that learns patterns from data.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: f1c5087e-477a-4b5f-85f7-1c419e103ddf.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-16T07:47:42.153Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 4e8c4cdc8fff0517f8836f26b5cd4ddfe7a035d136b7c2becf633a6f1336b952. Source text hash: 7babea7da29ea9f87ce80bb9c3921f0fc3611048b99c39bc69fd071e405e8081.
passed · 2026-11-16T07:47:42.153Z · anthropic/claude-opus-5.5
Publication: f1c5087e-477a-4b5f-85f7-1c419e103ddf
attribution — pass: The reading attributes the study to Zhaojia Wang, Jiucui Lu and Zheng Li, matching the arXiv listing. Findings are framed as what the authors report or argue, not as established fact.
Authors:Zhaojia Wang, Jiucui Lu, Zheng Li
summary — pass: The title matches the source, and the summary faithfully covers the research question, test setup, three findings and the authors' safeguarding argument. It glosses weight-space programming as directly changing a network's weights. Terms such as backdoor, MLP, MNIST and clean performance are left for tooltips. A small optional repair: add that the edits followed natural-language instructions without training data or gradients, which is the paper's central novelty. Its absence does not make the reading unclear. The question follows from the stated scope limits.
directly modify the weights of target models according to natural-language instructions
date — pass: originalPublished is 2026-11-12. This matches the v1 submission date on arXiv, not the later feed date of 2026-11-16.
[Submitted on 12 Nov 2026]
support — pass: Every claim is supported by an exact abstract excerpt. This includes six parameters and 0.02% of the weights, and the point that all five LLMs succeeded on the MLP but only GPT-6 Astra succeeded on all three targets. The closing sentence on scope is a cautious, bounded interpretation that does not overstate generality. The reading does not independently verify the paper's results.
including only six parameters for the MLP and 0.02% of the weights for a fine-tuned Qwen model
media — pass: The selected media is the verified arXiv abstract page for this paper, which is the direct original source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A language model with many learned numerical settings, used to process and generate language.
Definition referenceA system of connected simple calculations, loosely inspired by the brain, that learns patterns from data.
Definition referenceFurther training of an existing model to adapt it to a particular task or use.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA language model with many learned numerical settings, used to process and generate language.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceA system of connected simple calculations, loosely inspired by the brain, that learns patterns from data.
Definition referenceFurther training of an existing model to adapt it to a particular task or use.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA language model with many learned numerical settings, used to process and generate language.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition reference