Back to list
Product LaunchAI SecurityLarge Language ModelsCybersecurity

Zro Launches Uncensored Security Models Dolly 1 and Dolly 2 for Defensive Cybersecurity and Red Teaming

Product Hunt has featured the launch of Zro's new uncensored security models, Dolly 1 Security and Dolly 2 Security, developed to overcome standard LLM refusal mechanisms during authorized cybersecurity research. Built on GLM-5.3 Flash and GLM-5.3 architectures, these models utilize abliteration techniques to eliminate false-positive refusals while retaining foundational model capabilities. Designed specifically for defensive testing and red-team operations, Zro integrates these systems into an inference platform characterized by strict zero request retention and multi-region execution. With transparent API and CLI access, competitive token pricing, and open model evaluations, Zro provides security analysts and automated coding agents with the unhindered tools required to simulate threats and audit software vulnerabilities without storing proprietary prompts or sensitive codebases.

Product Hunt

Key Takeaways

  • Dedicated Security Models: Zro has released Dolly 1 Security and Dolly 2 Security, specialized models engineered specifically for defensive penetration testing, vulnerability discovery, and automated red-team workflows.
  • Overcoming Safety Refusals: By applying in-house abliteration methods to open architectures like GLM-5.3 and GLM-5.3 Flash, Zro removes automated refusal triggers that conventionally impede legitimate cybersecurity investigations.
  • Zero Request Retention Guarantee: Built to protect sensitive vulnerability and codebase intelligence, Zro operates across multi-region infrastructure without logging, caching, or training on user prompts or code.
  • Accessible Deployment Options: The launch introduces both API and Command-Line Interface (CLI) access, complemented by transparent pricing tiers and published evaluation weights for community verification.

In-Depth Analysis

Addressing the Refusal Dilemma in Cybersecurity Research

Modern large language models are engineered with extensive alignment training, safety guardrails, and reinforcement learning with human feedback (RLHF) designed to prevent malicious usage. However, these generalized safety constraints frequently trigger false positives during legitimate security workflows. When security researchers, red-team engineers, and penetration testers prompt models with reverse engineering tasks, vulnerability identification routines, or exploit reproduction payloads, general-purpose models routinely decline execution. This friction severely limits the productivity of defensive testing teams and prevents coding agents from evaluating software resilience at scale.

Zro’s introduction of Dolly 1 Security and Dolly 2 Security targets this operational gap. Rather than allowing refusal layers to interrupt critical research workflows, Zro provides abliterated model checkpoints tailored to execute technical security prompts completely. Security personnel can simulate attack vectors, evaluate defensive postures, and stress-test code repositories without constantly navigating superficial refusals that mistake defensive testing for illicit activity.

Technical Foundation: GLM Adaptation and Abliteration

The technological backbone of Zro's launch relies on adapted variants of advanced open foundation architectures, specifically GLM-5.3 Flash and GLM-5.3. Dolly 1 Security is designed as a lightweight, low-latency variant based on GLM-5.3 Flash, optimized for high-volume automated vulnerability scanning and rapid agent operations. Dolly 2 Security represents a larger-scale implementation based on the full GLM-5.3 architecture, subjected to an in-house abliteration procedure engineered to suppress refusal directions in weight activations while strictly preserving core reasoning, syntax analysis, and code synthesis capabilities.

Alongside the hosted availability of Dolly 2 Security, Zro has published model evaluation benchmarks and technical weights, providing transparency for independent verification. Commercial access is structured around competitive inference rates: Dolly 1 Security is priced at $0.45 per million input tokens and $1.50 per million output tokens, while Dolly 2 Security is available at $2.10 per million input tokens and $6.30 per million output tokens. A free tier is also offered for daily exploratory usage on Flash-grade models.

Zero-Retention Architecture and Agent Workflow Integration

In vulnerability assessment and exploit research, information privacy is paramount. Enterprise security teams and red teams cannot risk transmitting unpatched zero-day vulnerabilities, confidential software proprietary logic, or proprietary attack surface topologies to platforms that log inputs for downstream analysis or retraining. Zro tackles this compliance obstacle with an architectural commitment to zero request retention.

Under this security posture, neither code submissions nor prompt sequences are persisted on disk or ingested into subsequent training datasets. Through multi-region distributed execution, queries are processed statelessly in memory across high-throughput GPU infrastructure. The service natively supports developer integration through command-line utilities and standardized API endpoints compatible with libraries like LiteLLM. This architecture enables automated coding agents, continuous integration scanners, and penetration testing frameworks to trigger inference in a single command without creating external compliance liabilities.

Industry Impact

The introduction of uncensored, abliterated models purpose-built for cybersecurity represents an important shift in how the generative AI industry handles domain-specific capabilities versus horizontal alignment. As defensive security increasingly shifts toward autonomous agents, reliance on generic chatbots with restrictive guardrails is proving untenable. Security engineers require models that treat complex vulnerability research as an analytical discipline rather than a policy violation.

Furthermore, Zro's strategy highlights the growing viability of abliteration techniques—selectively removing refusal vectors while maintaining general capabilities—over full re-training from scratch. By decoupling privacy preservation and specialized model steering from proprietary closed-ecosystem platforms, this launch fosters an environment where defensive teams can run unhindered red-team simulations. However, it also reinforces ongoing industry discussions regarding the dual-use nature of advanced weights, putting greater responsibility on infrastructure layers to balance open research utility with accountable operational governance.

Frequently Asked Questions

What are Zro's Dolly 1 Security and Dolly 2 Security models?

Dolly 1 Security and Dolly 2 Security are uncensored AI models developed by Zro for cybersecurity professionals and red teams. Based on GLM-5.3 Flash and GLM-5.3 architectures, these models have undergone abliteration to remove standard refusal behavior, allowing defensive researchers to analyze vulnerabilities, conduct exploit testing, and automate security audits without being blocked by false-positive alignment triggers.

How does Zro address privacy concerns during security audits?

Zro utilizes a zero-retention framework that guarantees user prompts, vulnerability data, and source code are never stored, cached, or utilized for future model training. Inference is executed in memory across multi-region infrastructure, ensuring that proprietary software repositories and sensitive threat intelligence remain strictly confidential.

How can developers and security teams access these models?

Users can access Dolly 1 Security and Dolly 2 Security through the Zro API or dedicated CLI tools. The service provides standardized integration points for coding agents, supported by tiered per-token pricing alongside free daily allotments for testing on Flash-tier infrastructure.

Related News

Gemini API Tool Retrieval Preview Gives Agents Access to Hundreds of Tools With Fewer Tokens
Product Launch

Gemini API Tool Retrieval Preview Gives Agents Access to Hundreds of Tools With Fewer Tokens

Google AI Blog announced that Tool Retrieval is now in preview in the Gemini API. The feature is presented as a way to give agents access to hundreds of tools while using fewer tokens. The announcement pairs those two points: the breadth of tool access and lower token use. It does not explain how Tool Retrieval works, specify how token use was measured, or quantify the reduction. It also does not provide an exact tool count beyond describing access to hundreds of tools. The stated release status is preview, and the announcement does not give a timeline for general availability.

LangSmith Tool Gateway Enters Public Beta for Secure, Governed AI Agent Tool Use
Product Launch

LangSmith Tool Gateway Enters Public Beta for Secure, Governed AI Agent Tool Use

LangChain announced that LangSmith Tool Gateway is in Public Beta. The gateway handles credentials, identity, and access for every agent tool call, with policies, approvals, and traces. The announcement frames secure, governed tool use around the individual call: its focus is on how agents access tools, rather than on what those tools or agents can do. The update identifies what the gateway covers but does not explain how credentials are managed, how access decisions are made, when approvals apply, or what information traces contain. The Public Beta is the only release status stated in the announcement.

Google’s Gemini Decision 1 Picks Answers Instead of Writing Them Across Five Input Types
Product Launch

Google’s Gemini Decision 1 Picks Answers Instead of Writing Them Across Five Input Types

Google’s Gemini Decision 1 is a model designed to pick answers rather than write them, according to Shittu Olumide’s report for KDnuggets. It reads text, images, video, audio, and PDFs in a single forward pass and supports up to 262,144 input tokens per request. The central distinction is between what the model reads and what it does with that input: it accepts several kinds of material, but its stated task is answer selection, not answer writing. The token figure describes the maximum input per request, not how much the model outputs. The report does not explain how answers are presented for selection or provide results showing how well the model performs.