Zro Launches Uncensored Security Models: GLM-Based Dolly Series Powers Red Team Testing

Title

Zro Launches Uncensored Security Models: GLM-Based Dolly Series Powers Red Team Testing

Meta Description

Zro launches Dolly 1 Security and Dolly 2 Security on Product Hunt—two uncensored models based on GLM-5.3 Flash and GLM-5.3 with abliteration applied, offering zero request retention via API and CLI for defensive testing and red-team workflows.

Article Content

Security research often requires working with sensitive questions that general-purpose LLMs will simply refuse to answer. A project called Zro on Product Hunt aims to resolve this tension: it has released two "uncensored" model variants, Dolly 1 Security and Dolly 2 Security, based on GLM-5.3 Flash and GLM-5.3 respectively, with abliteration (ablation-based de-alignment) applied, targeting defensive testing and red-team workflows.

Why Security Teams Need "Uncensored" Models

General-purpose chat models almost universally ship with strict refusal policies. When security researchers need to analyze the intent behind a malicious sample, construct phishing email examples, or walk through an attack chain, the model often responds with "I can't help with that," bringing the workflow to a halt. Zro's approach is to "liberate" these capabilities from general models and dedicate them to authorized security testing.

Abliteration is a technique that weakens the refusal direction at the model weight level. It does not involve retraining; instead, it uses targeted interventions to reduce the model's tendency to trigger refusal behavior. Building on this, Zro produced two variants: Dolly 1 Security (based on GLM-5.3 Flash, more lightweight) and Dolly 2 Security (based on GLM-5.3, more capable).

Product Form: API and CLI, Zero Request Retention

Zro offers two access paths: the Zro API and a command-line tool (CLI). For red teams, the CLI makes it easy to integrate into automation scripts and local workflows, while the API suits batch calls and platform-level deployments.

You may not have noticed its privacy commitment: Zro states that it enforces "zero request retention." For work involving sensitive security content, this matters more than raw model capability—attack payloads, internal asset information, and vulnerability details should never persist on third-party servers.

Iteration Since Launch

According to the official description, since launching in July, Zro has progressively added new models, agent integration capabilities, an annual subscription plan, and a referral mechanism. This suggests the product is evolving from a single model interface into a more complete security research toolchain. Agent integration is particularly worth watching—many steps in security testing workflows require models to autonomously call tools and chain together multi-step operations.

A Balanced View of This Product Category's Boundaries

De-aligned models are not without cost. Weakening refusal behavior typically comes with a decline in general safety alignment, which may affect the model's reliability in non-security scenarios and raises the risk of misuse. Zro explicitly limits its positioning to "defensive testing and red-team workflows," but the compliance responsibility in actual use still falls on the user: whether authorization has been obtained, whether testing runs in an isolated environment, and whether outputs are properly controlled are all questions to evaluate before purchasing.

In addition, building secondary processing on top of a third-party open-source base (the GLM series) means the model's capability ceiling is constrained by that base. For high-intensity adversarial scenarios, teams still need to combine their own evaluations to determine whether it meets their needs, rather than making decisions based solely on the "uncensored" label.

Red Team Workflows and Authorization Boundaries

Red teaming refers to authorized testing activities that simulate attacks on one's own systems from an attacker's perspective, with the goal of finding weaknesses before real attackers do. Its workflow typically includes reconnaissance, constructing attack payloads (such as phishing emails or malicious script samples), walking through multi-step attack chains, and validating the effectiveness of detection rules. These steps require the model to generate content that is considered "harmful" in general contexts, and is therefore often refused by general models. Compliant red-team activities are premised on written authorization, a clearly defined scope, and an isolated environment; the resulting attack samples are sensitive assets that must be stored and transferred under controlled conditions. This also explains why zero request retention is so valuable to these users: test payloads themselves may contain exploitation details for unpatched vulnerabilities, and any leak would constitute a real risk. Users should confirm they are operating within an authorized scope and comply with the regulations of their jurisdiction regarding security testing and model use.

Who It's For

Based on its positioning, Zro's target users are penetration testing teams, security research labs, and defenders who need to simulate an attacker's perspective. For individual developers, the low barrier to entry via API and CLI also reduces the cost of trying it out. However, in the product's early stage, it is advisable to first validate capability and stability with non-sensitive samples before gradually introducing it into real workflows.

Background

What Are GLM-5.3 and GLM-5.3 Flash

The GLM series is an open-source large language model family from Zhipu AI, trained with a general language modeling objective and covering bilingual Chinese-English capabilities along with coding and reasoning. The GLM-5.3 and GLM-5.3 Flash mentioned in this article are two specifications within that family: the former is the more fully capable standard version with a larger parameter count and higher inference cost; the latter, named "Flash," typically refers to a distilled or structurally streamlined lightweight version with lower inference cost, suited to high-concurrency or resource-constrained deployments. Both come with the vendor's built-in safety alignment and refusal policies—which is precisely why Zro chose to apply abliteration on top of them: the base is capable enough, and only the refusal direction needs to be weakened to redirect it toward security testing. Understanding this helps clarify that the capability difference between Dolly 1 and Dolly 2 essentially stems from the difference in base specifications, not from differing levels of training investment by Zro itself.

The Technical Principles of Abliteration

Abliteration is a portmanteau of "ablation" and "obliteration," referring to removing a model's refusal behavior directly at the weight level without retraining. The typical approach is: construct a set of harmful instructions and a set of harmless instructions, run forward passes on each to extract activations from the model's layers (especially the residual stream), compute the difference between the two means to obtain a vector representing the "refusal direction"; then, at inference time, project that direction out of the activations of the relevant layers, or directly apply an orthogonalization correction to the weight matrices, so that the model no longer produces refusals along that direction. Because the entire process requires no gradient updates or labeled data, it costs far less than RLHF re-alignment and can be completed in anywhere from tens of minutes to a few hours. The trade-off is that the refusal direction often shares part of its subspace with general safety and cautious expression; once erased, the model may also become overly compliant or less discerning in non-attack scenarios.

One-line Summary

Zro launches Dolly security models de-aligned via abliteration on the GLM series, serving red-team testing.

Paragraph Summary

Zro has released two "uncensored" model variants on Product Hunt—Dolly 1 Security and Dolly 2 Security—based on GLM-5.3 Flash and GLM-5.3 respectively, using abliteration to weaken refusal behavior at the weight level for defensive testing and red-team workflows. The product offers API and CLI access and promises zero request retention, making it suitable for handling sensitive content such as attack payloads and vulnerability details. Since launching in July, Zro has added new models, agent integration, annual subscriptions, and a referral mechanism, evolving from a single model interface into a security research toolchain. The article also cautions that de-alignment weakens general safety alignment and raises misuse risk, that the capability ceiling is constrained by the third-party base, and that compliance and authorization responsibilities remain with the user.