Security1 distinct publisher2 min readPublished
Open weights could always be modified. Hosting them for anyone who signs up is the new part. That lands the uncensored-model question on whoever approves red-team tooling. Whether that helps defenders is contested.
The Watch · Security desk

Compiled by The WatchSomething wrong?How this is made
The weights are open, so removing refusals from them was already available to anyone willing to do the work. Abliteration.ai supplies the part that took effort: hosted builds of models including GLM-5.3, reachable from a web browser or an API, according to the TechCrunch report summarized by SC Media [1][2].
The report covers what the service does and how it's reached, but it stops short of price, launch date, customer count, capability measurement, or any documented incident tied to the platform [11]. That shapes how the story reads. The purchasing question is live now. The record so far says little about what the service actually produces. A red team that asks to buy access has the vendor's stated purpose [3] and very little else to put in a risk memo.
Co-founder Devon's argument is about tempo: defenders model bad actors so they can move as fast as possible [4]. Andrew Yoon's objection is a forecast about who ends up using the output [9]. Both are claims about future behavior, and nothing in the brief measures either one.
What is checkable is the control surface. The moderation layer is something customers implement for themselves [5], which governs what a buyer's own agents are allowed to do rather than which buyers exist. Violence prevention is described as work in progress [6]. Both sit downstream of the account. All three controls take effect only after signup, so the account holder's identity is not among the things they decide [8]. For a buyer that means the assurance on offer is the vendor's word about its own roster, and for anyone else it means the endpoint answers the same way regardless of which side of an engagement the query comes from [7].
That is also the practical difference between renting this and running your own. An API service means the prompts, and whatever target detail rides in them, leave the buyer's environment [2]. The teams that fine-tune models they already operate are buying nothing new; they are keeping the transcript [10]. Whoever signs off on offensive tooling now owns that trade. The brief leaves them one fact: identity verification has not been implemented [7].
Ranked by verification strength, evidence, and original report placement.
TechCrunch reports that Abliteration.ai has launched a service providing access to open-weight AI models with their safety guardrails removed, making it easier for users to perform tasks that standard AI models would refuse.
The platform offers modified versions of models like GLM-5.3, allowing queries via a web browser or API.
Abliteration.ai aims to enable offensive cyber operations, red-teaming, and agent testing that other models refuse.
Co-founder Devon stated that the goal is to allow defenders to "move as fast as possible" by modeling bad actors.
Abliteration.ai offers a moderation layer for users to implement their own guardrails.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · September 4, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
product
Abliteration.ai sells hosted access to a GLM-5.3 with its refusals removed1 distinct publisher
science
GLM-5.3 says the quiet part: the base model did not change, the post-training did1 distinct publisher
build
AMD and NVIDIA top Hugging Face's new-model count with converted checkpoints1 distinct publisher
product
Cheap bug-hunting arrives: GLM 5.3 puts near-frontier vulnerability discovery on your own hardware1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Single secondhand brief
Everything traces to one SC Media brief that is itself summarising a TechCrunch story, so the whole account is a retelling of someone else's interview. No company page, pricing sheet, model card, or licence appears anywhere in it. Devon's rationale and Yoon's sociopath comparison are quoted rather than tested, and nothing the service produced has been examined.
Launch only, no users
The one datable event is the launch itself: access paths exist and a modified GLM-5.3 is named. Nobody who uses the service is identified, no query or seat volume is disclosed, and the security firms mentioned are weighing the category rather than named as buyers of this product.
Pitch ahead of proof
The "move as fast as possible" case for defenders and the warning about sociopathic models both describe capability nobody in this reporting has measured. The parts that are concrete point the other way: filtering is the buyer's job, violence controls are unfinished, and the account holder is unverified. Both the promise and the alarm outrun what has been shown.
Seller and trade venue
The claim that hosted uncensored models help defenders comes from a co-founder selling access to them. The counterweight is a safety expert whose affiliation the brief never gives, the split among cybersecurity firms is attributed to nobody by name, and the venue is a security trade outlet that ends by pitching its own AI guidance. No disinterested voice appears in the story.
Consistent but unconfirmed
The facts are few and they hang together: a hosted service, one named model, two access paths, three controls in varying states of completeness, and SC Media does not overclaim on its own account. What holds confidence down is that a single secondhand summary carries all of it, and the detail a buyer would act on, the missing identity verification, has no second confirmation.