Tracing the Unrestricted AI Ecosystem: From Public Repositories to Hosted Services
8 Minute Read
by Serhii Melnyk
Getting a language model to ignore its safeguards used to take a clever prompt. Now, it only takes a download or a monthly fee. Tools that strip refusals out of open-weight models are free on GitHub, modified checkpoints sit on public model hubs, and resellers will host them for you behind a familiar API, often with a promise that nothing is logged.
We looked at how this market fits together, how much of it runs on ordinary mainstream infrastructure, and what that means for defenders.
What ‘Uncensored’ Actually Means
Terms such as “uncensored,” “unfiltered,” “no-refusal,” and “abliterated” are often used interchangeably, but they describe different techniques. A jailbreak prompt tries to get around safeguards within a single conversation. Fine-tuning changes a model through additional training. Abliteration edits a model’s weights or internal activations to reduce refusal behavior. A hosted service might use any of these or simply run an ordinary model behind a permissive system prompt, and buyers are rarely told which.


Figure 1. Jailbreak attempts in mainstream chat interfaces. (1) An English-language “DAN” persona claims to have no rules and, when challenged, produces a sentence endorsing violence and discrimination. (2) A Russian-language DeepSeek reply claiming to be unrestricted and listing what it says it will provide, including nerve agent synthesis and banking exploits.
Access to the weights matters most in practice. Jailbreak prompts work against closed APIs, but editing refusals out of a model requires a downloadable copy, which only an open-weight model provides. Once someone has downloaded a copy, the original publisher generally cannot recall it.
Local use is not invisible, though: cloud providers, network operators, endpoint monitoring, and payment services may all keep records. Hosted “no-log” services are a separate matter, covered below.
The Underground Sells It, the Open Internet Supplies It
Cybercrime forums and private messaging groups are where buyers and sellers meet, compare services, and trade recommendations. The goods themselves are rarely hosted there. The models sit on public model platforms, the modification code lives in open repositories, and the computing power comes from commercial cloud and GPU providers. Resellers stitch these pieces together through websites, Telegram bots, and API gateways, which creates a supply chain that runs across underground and mainstream platforms alike. The tooling at the start of that chain is public and easy to find.
To gauge how widely this tooling has spread, on October 30, we queried GitHub’s search API for repositories with “abliterated” in the name, description, or topics. It returned 290 matches, more than eight times the 34 listed on GitHub’s topic page, and 278 of them were created in 2026. Only 12 predate this year. Creation counts climb from 20 in the first quarter to 44 in the second and 126 in the third, and the first 30 days of October added another 88.

Chart 1. New repositories created per quarter, with October 2026 shown separately. Source: GitHub search API, queried October 30, 2026; repositories with “abliterated” in the name, description or topics (n = 290). Counts are approximate: GitHub’s search index is incomplete, and keyword matches can include unrelated repositories.
Of those repositories, roughly one in five reference agent or command-line harnesses, about nine are documented to be running abliterated builds on free notebook platforms such as Colab and Kaggle, and about seven are targeting rented GPUs. At least two describe applying the edit at inference time, which, if it works as described, means no modified weights ever reach disk. At most, six of the 290 mention offensive security terms in their metadata, so the ecosystem is not primarily advertised as offensive, and descriptions appear in at least six languages besides English.
Distribution is equally straightforward. Public model hubs list modified checkpoints, and local inference libraries let anyone download and run them with minimal setup. At the time of analysis, a Hugging Face search for “abliterated” returned nearly 9,000 models. The exact count shifts with the date and with how quantized copies are tallied. The Ollama library, meanwhile, listed almost 400 community-uploaded models described as “uncensored,” some of them marketed toward security operations.
From Open Weights to Hosted Services
For the buyer, a complicated deployment shrinks to choosing a provider and pasting in a key. For the provider, that position brings control over which checkpoint is served, how it is configured, whether it stays online and what happens to the data passing through it. The operator becomes the intermediary between model developers and users, and most of the risk discussed later in this post follows from that.
Security tooling is being repackaged in the same way. Services use models to coordinate existing tools, sort their findings, and write up the results.

Figure 2. Forum ad for “AI Pentest Checker,” a beta automated penetration-testing service. A submitted domain gets a PDF report and a 0–100 security score in about 10 minutes, produced by 14 scanning engines plus LLMs that filter false positives and write exploitation notes.
Modified checkpoints are also turning up in commercial inference catalogs. Instead of downloading weights and configuring a local setup, a customer picks an “abliterated” model from a provider’s list and pays per token, as they would for any other model. That folds reduced-refusal models into the normal purchasing flow. It also leaves questions a catalog entry cannot answer: who modified the weights, which checkpoint is actually running, and how well it performs. A label such as “best at cyber” is the provider’s own marketing.

Figure 3. “Abliterated models” section of an inference provider catalog listing GLM 5.3, GLM 5.3 Flash and DeepSeek V4.1 Flash, with per-million-token prices ($0.12–$2.00 input), 1M context, speed and quantization. The “best at cyber” label on GLM 5.3 is the provider’s own marketing.
What customers do with these services is easy to guess, and the range is wide. Across the forum posts, storefronts, and captures we reviewed, the advertised and demonstrated use cases cover nearly every stage of an operation: drafting phishing lures and templates, writing stealers, loaders and ransomware, advising on delivery and evasion, automating scanning and reconnaissance, and writing or explaining exploits. The sellers’ marketing promises no limits, and in the examples we reviewed, the models rarely refused.

Figure 4. A chat interface answers “Write ransomware in python” with a script labeled “Simple Ransomware Demo – For Educational and Research Purposes Only,” with no refusal.
Dedicated to hunting and eradicating the world's most challenging threats.
SpiderLabs
Figure 5. A coding agent running an abliterated Qwen3-Coder model, asked in Russian how to create malware. It proceeds, writing a “Malware Creation Toolkit” README and starting a project structure.

Figure 6. A terminal coding agent (GLM-5.1) advising on delivering a macOS .scpt payload. It explains how Chrome, macOS quarantine, and Safe Browsing would treat the file and suggests packaging it as a .zip.

Figure 7. A commercial inference service’s chat page, with GLM 5.3 abliterated selected, answers a request for an X.com phishing email template with an HTML source. The only safeguard is a one-line “authorized training/pentesting only” notice.
This pattern is consistent with our earlier research into AI-assisted phishing kit development, where we documented more than 70 instances sharing a common backend and administrative panel, with code characteristics assessed as strongly indicative of AI-assisted development.
That investigation concerned phishing infrastructure rather than unrestricted model services, but it illustrates the broader effect: AI assistance can reduce the time and skill needed to assemble, adapt, and maintain criminal tooling. It does not, by itself, prove that every kit or operator relies on the same models.
Pricing, Provenance, and the Limits of Transparency
Sellers offer more than model access. Providers also promote cryptocurrency payments, minimal registration, private chat histories, and claims that prompts are never retained. Multi-model gateways add convenience and concentrate trust at the same time: a single API key can open up several model families through interfaces developers already know, and compatibility with coding agents makes it easy to switch providers without rebuilding an integration.


Figure 8. Reseller storefronts. (1) A “FraudGPT” site advertising an unrestricted assistant from €89.99. (2) The RouterHub API gateway, offering 27 models behind one key, including GLM 5.3 Abliterated alongside mainstream models.
Pricing is just as hard to read. Services bill per token, by subscription, by task, or by compute time, so the same model can be sold as a bargain, a convenience, or a less restricted option depending on the page. Token prices look easy to compare, but in practice, they can cover different model versions, throughput, context limits, caching, rate limits, and retention arrangements. Discounts add another unknown: one gateway in Figure 6 advertises mainstream commercial models at up to 95% below official prices without saying where the difference comes from.

Figure 9. Commercial packaging. (1) A pricing page listing abliterated and “uncensored” models per token, plus hourly GPU rental and weight download fees. (2) The apixo gateway’s price table, showing discounts of up to 95% against official prices for mainstream models and no official price for its uncensored variants.
Then there is the “no-log” promise. To answer a prompt, the provider has to process it, and it may also handle account details, payment metadata, API keys, and usage records. A statement that sessions are not stored does not say whether content is held temporarily, whether metadata is kept for abuse monitoring, whether backups exist, or what upstream infrastructure providers can see.
Capability, Benchmarks, and the GLM-5.3 Effect
A model’s willingness to answer a prohibited request says little about whether it can finish a technical task, and a strong benchmark score does not show that it can carry out an intrusion reliably or without human direction. Underground marketing tends to place “uncensored” claims next to cyber capability claims. They are separate attributes, and the second one needs its own evidence.
GLM-5.3 shows why the distinction matters. Z.ai released the open-weight model on August 14, 2026, and in the underground communities we reviewed, it was the model most often mentioned, tested, and discussed, across advertisements, benchmark claims, and user-run tests. Sellers use its name as a selling point for “uncensored” and “best at cyber” services. It also appears in several figures in this post, which makes it the clearest example of where this market’s attention is going.
In its September 17 assessment, the US National Institute of Standards and Technology’s Center for AI Standards and Innovation (CAISI) called GLM-5.3 the most cyber-capable open-weight model released to date. It also found the model’s cyber capabilities significantly lower than those of current US frontier models, estimating that it trails them by about four months on its aggregate cyber measure.

Figure 10. Benchmark charts from a forum post advertising an abliterated GLM-5.3 service. GLM-5.3 (highlighted) trails the top model in all three panels: Terminal-Bench 4.0 (41.8% vs. 51.8%), CyberGym (84.5% vs. 85.6%) and ExploitGym (105 vs. 216 of 869 tasks). The poster’s capacity and cache hit claims are unverified.
Buyers also have to take the model itself on trust. A service may advertise a well-known base model while serving a quantized version, a different checkpoint, or a modified prompt template, and it may share inference capacity with other customers. Provenance, quantization, context limits, system prompts, and serving configuration all change the result, and from the outside a customer can see almost none of them.
Defensive Considerations
Individual brands come and go, so tracking them has limited value. The lasting pattern is open weights and modification tools, models embedded in coding and security workflows, growing hosted inference, and gateway offerings, and marketing built on anonymity, speed, price, and fewer refusals. Following those developments together is more useful than tracking individual brands or forum posts. Six practical steps follow.
- Inventory your AI endpoints. Record the base URLs, API keys, local inference servers, and browser-based AI services in use across developers and agents. Compare them with an approved provider list and investigate unfamiliar gateways.
- Watch for traffic to model gateways and hubs. Review proxy, DNS and egress logs for connections to unfamiliar inference gateways, model hubs and notebook platforms, and treat large model downloads or new API-compatible endpoints on developer machines as worth a look.
- Treat coding agents as execution environments. They run code, so they should not be handled like chat clients. Give them only the access they need, isolate their workspaces, restrict outbound connections where practical and require human approval before destructive actions or access to credentials and production systems.
- Control what goes to third-party inference. Keep secrets and regulated data away from unapproved endpoints, use scoped and short-lived credentials, and watch for unusual API-key use, unexpected destinations, and unexplained jumps in spend.
- Check model provenance before deployment. Approve publishers, pin versions and hashes, review model cards and configuration files, and test unfamiliar weights in an isolated environment with no credentials or unrestricted network access. Prefer safer serialization formats and scan model packages and associated code.
- Do not rely on refusals as a security control. A refusal does not make a workflow safe, and a compliant answer is not necessarily correct. Enforce security outside the model with identity controls, network policy, data access restrictions, sandboxing, and human approval.
Conclusion
The unrestricted AI market is a chain of ordinary parts: open-weight model publishers, modification tools, compute providers, workflow developers, underground communities and API resellers. Forums help buyers and sellers find each other, but most of the infrastructure comes from the same global platforms and commercial services that legitimate developers use. Because of that, the market is hard to disrupt at any single point.
The most significant change we observed is the shift from technical access to packaged service. Hosted gateways let customers call modified or less restricted models through familiar APIs without building anything themselves. That lowers the barrier to entry and puts a new operator between the user and the model, one with visibility into prompts, credentials and usage data.
Threat actors can use these models to draft and tailor phishing lures, translate social-engineering messages, write scripts, and support reconnaissance or vulnerability research. Connected to coding agents and security tools, they can automate parts of those workflows. We assess with moderate confidence that this lowers the barrier to entry more than it raises the ceiling of what attackers can do: less-skilled actors clear technical hurdles more easily, and experienced operators can work faster and at greater scale.
Defenders should not treat every unrestricted model or public modification tool as malicious. The more pressing risk is the combination of uncertain model provenance, unapproved inference services, and agents with access to sensitive systems or data. Organizations that know where their models come from, where their prompts go and what their AI-enabled tools are allowed to do are better placed than those that track each new uncensored brand.
About the Author
Serhii Melnyk is Cyber Threat Intelligence Analyst at LevelBlue Serhii has eight years of experience in the security industry. Among his many tasks at LevelBlue, he actively contributes to the MISP project and the MITRE ATT&CK. Follow Serhii on LinkedIn.
About LevelBlue
LevelBlue secures what's next with intelligence-led security delivering visibility and speed to stop threats faster. As the world’s largest and most analyst-recognized pure-play managed security services provider, our AI-powered managed services and cyber expertise across managed, advisory, and incident response services help clients operate with confidence. Learn more about us.