Open-Weight AI Models Without Safety Filters, Now as a Service
Removing safety guardrails from AI models has long been a practice carried out by researchers and developers in open-source communities. A startup has now transformed that underground activity into a paid service, hosting modified versions of open-weight models that no longer refuse harmful requests.
The platform allows users to query these uncensored models through a web browser or via an API, eliminating the technical barriers that previously required users to download pre-modified models and secure their own computing infrastructure. Among the models available is a recently released open-weight system that has been stripped of its refusal mechanisms.
The Cybersecurity Argument for Unfiltered AI
The company’s stated mission is to enable offensive cybersecurity work, red-team exercises, and agent testing that other models refuse to support. The reasoning mirrors a well-established principle in security research: you cannot defend against a threat you cannot reproduce. A model that refuses to generate working exploit code cannot help a red team prepare for real attackers.
The startup’s co-founder argues that uncensored frontier models serve as the best form of defense. « The big picture is that abliterated models can model bad actors, » he said, explaining that defenders need tools capable of simulating adversarial behavior at the same speed attackers operate.
According to the company, its customers include early-stage red-teaming startups based in the United Kingdom and Europe. These firms help banks, airlines, and enterprises managing critical infrastructure strengthen their cybersecurity posture. The co-founder noted that one major customer red-teams AI agents used by banks, but cannot effectively do so with standard off-the-shelf models.
Easy Access, Minimal Oversight
Reporters were able to quickly create an account and begin querying an abliterated version of the model at no cost through a web browser. When asked to produce a Python program capable of stealing saved browser passwords and a detailed protocol for culturing a dangerous pathogen at home, the model complied readily in both cases.
The platform does include a basic moderation layer, and the company says it is working to add additional safeguards against violent content. During testing, the model refused to provide instructions related to self-harm. However, the company has implemented no identity verification procedures beyond logging the credit card used to access the service, acknowledging that determining who should receive access remains an unresolved challenge.
« You don’t want to be the person responsible for someone doing something crazy, so where do you draw the line of your responsibility as a company? » the co-founder said. « We’re still defining that. »
Industry Skepticism and Safety Concerns
Not everyone in the cybersecurity community agrees that abliterated models are essential tools. Several agent red-teaming companies say they rely instead on fine-tuning open-weight models that already have minimal guardrails, arguing that the abliteration process itself can degrade a model’s knowledge and capabilities.
One red-teaming CEO stated that using abliterated models for cyber or biological harm would actually be less effective than using fine-tuned alternatives. Others in the field acknowledged that while abliterated models might elicit useful behavior for stress-testing, they are not yet part of their daily workflows.
Safety advocates warn that making abliterated models available at scale could lead to genuine harm. An AI safety researcher described the practice as modifying a model so that « it becomes a sociopath, » noting that any input could receive compliance. The researcher expects edited, abliterated models to be used for harmful purposes in the near future.
Where Regulation Might Step In
While many experts believe the spread of abliterated models cannot be stopped, there is growing discussion about where governments can intervene. Proposed measures include requiring providers to run classifiers that detect and block harmful cyber and biological weapons activity, and mandating that companies renting direct access to advanced GPUs verify customer identities and deny access where dangerous misuse is suspected.
The startup has not raised venture capital and says it is funded entirely through customer revenue, with deals already in place with major cloud providers. It is currently in discussions with investors.
As increasingly capable models are released with downloadable weights, the question facing both industry and policymakers is clear: if anyone can strip a model’s safeguards, does making the resulting version easier to access strengthen defense or amplify danger?
