In a paper accepted at the Internet Measurement Conference (IMC 2026), four researchers led by Zilong Lin reveal how large language models (LLMs) with their safety guardrails removed are circulating. The team surveyed five model distribution platforms, including Hugging Face, and identified 22,096 "uncensored LLMs", models whose safeguards have been removed or weakened, or that never had any. According to lead author Lin, only about 13% of these models carried a warning label.

The survey highlights not only the techniques for removing safeguards, but also a deeper problem: risk information gets lost as modified models are copied and redistributed. The EU's AI regulation, the AI Act, requires providers of general-purpose AI models to meet obligations such as documentation. However, those obligations are unlikely to reach developers who modify an existing model with a small amount of compute, or those who redistribute the modified versions, as providers. In an estimate that uses the EU's guideline figure for the original model's training compute, removing the safeguards from a 4-billion-parameter-class model took about one fifty-thousandth of the compute threshold used to decide whether a modifier counts as a new provider.

AD

Over 20,000 models with safeguards removed, only about 13% carry warnings

Lin is an assistant professor at the University of Missouri–Kansas City. According to his research summary published on November 17, 2026, the data cutoff was March 1, 2026, and the survey covered five platforms: Hugging Face, OpenCSG, ModelScope, Kaggle, and John Snow Labs. The 22,096 uncensored LLMs were published by more than 2,100 developers, and 681 new ones appeared in February 2026 alone.

About 70% of the models stated in their public metadata that they had removed safety restrictions. More than 11,000 advertised capabilities tied to harmful uses, such as generating malicious code or offensive cybersecurity work. Usage is not small either: 67 models have been downloaded more than one million times.

By contrast, only about 13% carried a moderation warning. Moreover, threat information about the models was not consistently carried over when they were copied to different platforms. Lin points out that the models themselves move more easily than the information needed to assess their risks. The findings show a structure in which, as models with removed safeguards are redistributed, the information users need to understand the risks is lost.

Removing safeguards from a 4B model takes about 1/50,000 of the EU's threshold compute

A widely used technique for removing LLM safeguards is "abliteration," which strips out the internal representation tied to a model's refusal responses. Research by Arditi et al. examined 13 publicly available chat models and showed that the behavior of refusing harmful instructions is strongly tied to a specific direction inside the model. When the component along this direction is removed from the weights, the model responds even to requests it previously refused. Such modification requires almost no conventional additional training.

Tools that automate this kind of modification are now publicly available. Heretic, developed by Philipp Emanuel Weidmann, is one example. According to its developer, with default settings on an NVIDIA GeForce RTX 3090, it can process the 4-billion-parameter-class chat model "Qwen3-4B-Instruct-2507" in roughly 30 to 40 minutes. Heretic's default configuration uses a proprietary method called "Arbitrary-Rank Ablation (ARA)."

Meanwhile, the EU AI Act uses the training compute of a general-purpose AI model as a factor in deciding what is regulated. The European Commission's guidelines and FAQ treat training compute exceeding 10^23 FLOP (floating-point operations) as one indicator that a model qualifies as a general-purpose AI model. Whether a developer who modifies an existing general-purpose AI model takes on obligations as the "provider" of a new model depends on whether the modification significantly changes the model's generality, capabilities, or systemic risk.

The concrete benchmark offered is whether the training compute used for the modification exceeds one third of the original model's training compute. The Commission explains that modifications reaching this level are exceptional, and that most ordinary fine-tuning and small-scale modifications do not create new provider obligations. However, compute is an important indicator, not a uniform legal threshold under which obligations are always avoided. The guidelines themselves are not legally binding.

There are also certain exceptions for providers of the original model. Article 53(2) of Regulation (EU) 2024/1689 exempts general-purpose AI models released under open-source licenses with publicly available weights from obligations to draw up technical documentation for authorities and to provide information to downstream providers. The obligations to put in place a copyright policy and to publish a summary of training data remain. The exemption also does not apply to general-purpose AI models with systemic risk.

Comparing the compute needed to remove safeguards with the Commission's benchmark shows a very large gap. According to NVIDIA's whitepaper on the GA102 architecture, the theoretical peak FP16 Tensor performance of the RTX 3090 is 142 TFLOPS for dense operations and 284 TFLOPS for sparse operations. One TFLOPS equals one trillion floating-point operations per second.

Take the longer of Heretic's processing times, 40 minutes (2,400 seconds), and assume the GPU sustained the sparse theoretical peak of 284 TFLOPS the entire time. The total comes to about 6.8×10^17 FLOP. If the original model's training compute were exactly the 10^23 FLOP indicator for general-purpose AI models, the benchmark for treating the modifier as a provider would be one third of that, about 3.3×10^22 FLOP, and the compute used to remove the safeguards would be only about one fifty-thousandth of it. Using the dense figure of 142 TFLOPS, it would be about one ninety-eight-thousandth.

This estimate comes with three caveats. First, abliteration is mainly a matter of forward passes and weight projection, and it is unclear how far those operations count as "training compute" under the AI Act. Second, the Commission's guidelines and FAQ are not legally binding interpretations. Third, if the original model's actual training compute exceeds 10^23 FLOP, the compute benchmark for treating a modifier as a provider rises as well, and the gap widens further. Note also that the operation count here is a rough estimate that applies the GPU's theoretical peak performance across the whole processing time, not a measurement of actual compute consumed.

Furthermore, redistributing a modified model is lighter work than removing its safeguards. Many models published in formats such as GGUF and GPTQ are trained weights that have been quantized to reduce file size. Such conversion normally requires no additional training, so a redistributor is even less likely to meet the compute-based criteria for a modifier.

AD

About 70% of uncensored models on Hugging Face are redistributed copies

The circulation of uncensored models is not sustained only by the developers who remove the safeguards. Once a model is published, other accounts quantize the weights, convert them into formats that run easily even on personal PCs, and republish them. Quantization is a technique that represents weights with fewer bits, reducing the model's size and the memory required at runtime.

A preprint released by 10a Labs in September 2026 examines this redistribution. Of the 11,635 uncensored-model repositories the company identified on Hugging Face between January 2024 and March 2026, 8,164, about 70%, were redistributed versions compressed through quantization or similar methods. There were 3,471 original models, with an average of 2.4 redistributed versions made per original. Moreover, just three accounts published 52% of the redistributed versions.

The analysis used GPT-5 to classify the models, and did not count merged models combining several models or copies on other distribution platforms. One of the top three accounts was removed from Hugging Face after the survey period ended, but according to the paper, the 936 redistributed versions that account had created remained as downstream copies. The case shows that deleting the original distributing account does not make already-replicated models disappear.

Replicated models spread across platforms, but warning labels are not necessarily carried over. According to the abstract of the paper accepted at IMC 2026, 11,094 models formed 4,457 clusters that were replicated across multiple platforms. The main source of replication was Hugging Face, but moderation labels showed inconsistencies between platforms.

On Hugging Face, setting a not-for-all-audiences tag on a model card displays a notice. However, setting the tag is mainly left to the model's owner, though the platform may also add it as part of moderation. The company's content policy prohibits content related to creating malicious software and similar activity, but it also lists asking owners to cooperate or make corrections as a response when a violation is reported.

Lin told The Register that a model's safety is hard to judge just by examining its files, and that actually running the model to verify its responses to harmful requests takes computing resources and money. Because it is not easy for downstream users to verify individual models, warning labels are an important clue for understanding risk.

Uncensored models are already being built into real applications. In 10a Labs' survey, among READMEs of GitHub apps that use uncensored models, 43% referenced Ollama and 14% referenced direct downloads from Hugging Face. Management at the distribution-platform level also bears on how the models are later built into and used in apps.

If local execution can't be stopped, does managing redistribution still matter?

A strong objection to regulating model redistribution is that it cannot stop people from running an already downloaded model on a personal PC. Philipp Emanuel Weidmann, who developed a safeguard-removal tool, told NPR he opposes a situation in which a handful of large companies decide the scope of AI use. According to NPR, he is working to keep his tool available even if uncensored models are removed from distribution platforms.

In fact, removing redistribution hubs does not completely halt circulation. The 936 redistributed versions that 10a Labs identified remained as copies even after the original account was deleted. In July 2024, the US Department of Commerce's National Telecommunications and Information Administration (NTIA) also recommended that, rather than imposing immediate restrictions on publishing the weights of large AI models, authorities collect evidence on the risks and keep evaluating them.

However, being unable to fully recall models already in circulation is a separate matter from being unable to manage future redistribution. 10a Labs focused on the fact that more than half of redistributed versions were concentrated in just three accounts, and it calls this layer the "narrowest point" in model circulation. Acting against a small number of redistribution hubs could affect downstream availability more than comparable measures against the developers of original models, it says. The paper analyzes that halting publication of original models does little to reduce availability, whereas addressing the accounts that handle redistribution leaves room to curb circulation.

Still, as the 936 surviving copies show, managing redistribution cannot recall every existing model. What can realistically be addressed is the number of copies made from now on, and how warning labels are handled when models are redistributed. If risk information is properly attached to new copies, users at least gain more to go on when choosing a model.

In Japan, there have been cases in which individual users faced criminal liability over programs created by misusing generative AI. In May 2024, Tokyo's Metropolitan Police Department arrested a man in Kawasaki for creating a program resembling ransomware by misusing conversational generative AI, and in October of that year he received a three-year prison sentence, suspended for four years. In June 2026, Chiba Prefectural Police also arrested a 19-year-old in Otsu on suspicion of creating ransomware with AI. Both were criminal proceedings against the individuals who created the programs in question, not against the distribution channels of AI models themselves.

The institutional focus going forward is how the European Commission treats modifiers and redistributors of existing models. If the criteria are revised so that modifications that remove safeguards, and redistribution of such modified versions, carry provider obligations regardless of compute, obligations such as documentation could reach distribution channels that are hard to capture under the current system. Under the EU AI Act, providers of general-purpose AI models placed on the market before August 2, 2025 have until August 2, 2027 to comply, which will be one milestone for judging how the system operates.

Meanwhile, the effectiveness of warning labels can be evaluated from actual circulation data. If the share of models with warnings rose from about 13% today to a majority but downloads and app integrations did not fall, that would mean the labels had only a limited effect in curbing circulation. Going forward, the prevalence of warning labels and changes in redistribution, along with how the European Commission handles the modifier criteria, will serve as concrete indicators for evaluating measures to manage model circulation.