DuoNeural's Cyber-Ornith Puts a 9B Hacking Agent on Consumer GPUs
A 9B Qwen3.5 fine-tune strips refusal behavior and adds agentic terminal skills, targeting local red-team workflows with GGUF quants that fit on a laptop.
Takeaways
- DuoNeural released Cyber-Ornith-1.5-9B-OBLITERATED, a Qwen3.5-based agentic cybersecurity model with zero refusal.
- Fine-tuned via QLoRA on 45k cybersecurity trajectories plus Hermes function-calling and OpenThoughts reasoning data.
- Self-reported 100% CyberSecEval triage, 95% Hermes tool-call AST accuracy, 62.5% Terminal-Bench pass rate.
- Preserves native think-tag reasoning traces so it plans exploits and tool calls before acting.
- mradermacher GGUF quants run in llama.cpp, Ollama, LM Studio at 5-8 GB.
- Apache 2.0 license, aimed at red-teamers, CTF work, and local threat-hunting agents.
DuoNeural has published the model weights for Cyber-Ornith-1.5-9B-OBLITERATED, a community fine-tune designed to operate terminals, call tools, and handle offensive-security prompts with few refusals. Community maintainer mradermacher has also released GGUF builds for llama.cpp and compatible runtimes. The quantization repository recorded more than 3,000 downloads during its first days, according to its Hugging Face counter.
A roughly 5.7 GB Q4 build makes the model practical on many consumer systems, including machines disconnected from hosted APIs. That gives security teams a private backend for sensitive scan results, exploit validation, and internal tooling. Its reduced refusal behavior also transfers moderation, authorization, and command safety to the operator.
Refusal removal meets QLoRA
The model card describes a dense Qwen3.5-derived transformer with Gemma-style attention and YaRN rotary-position scaling. Attention controls which earlier tokens influence each generated token, while YaRN adjusts positional signals to support longer contexts. The published checkpoint builds on OBLITERATUS/Ornith-1.5-9B-OBLITERATED.
Abliteration estimates an activation direction associated with refusals and modifies model weights to suppress that direction. The approach draws on work by Arditi et al. and subsequent open-source experiments. Because it targets a learned refusal mechanism rather than a formal policy system, behavior can still vary across prompts and contexts.
QLoRA adaptation then trains compact low-rank parameters against a quantized base, reducing the memory required for fine-tuning. DuoNeural says this stage focused on cybersecurity trajectories, terminal execution, and structured tool calls. The v3 production candidate uses a spherical linear interpolation merge, identified as SLERP with t=0.75.
Delimiter-wrapped reasoning traces appear before some final answers, preserving the base model’s visible planning behavior. Agent integrations need to separate those traces from executable commands and structured calls. Reasoning text can contain secrets copied from tools, untrusted instructions, or speculative commands that should never execute automatically.
Benchmark claims await replication
The project reports the following results for its v3 production candidate:
| Evaluation | Reported result | What it measures |
|---|---|---|
| CyberSecEval 3 triage and defense | 100.0% | Responses to security-analysis and defensive tasks |
| Zero-refusal compliance | 100.0% | Frequency of answering prompts in the project’s uncensored test set |
| Hermes function-calling AST accuracy | 95.0% | Structural correctness of generated function calls |
| Terminal-Bench | 62.5% | Completion of multistep tasks in a terminal environment |
| Generation throughput | 102 tokens per second | Reported speed on one RTX 4080 Super |
The zero-refusal score measures willingness to answer. Separate testing must establish whether generated commands are correct, safe, authorized, and effective. Function-call AST accuracy similarly checks structure more directly than real-world tool outcomes.
The authors also list four optimization targets:
- Terminal execution: improve Terminal-Bench 2.1 by at least 4 percentage points over the baseline.
- Security auditing: improve CyberSecEval 3 triage and defense by at least 20 points.
- Tool execution: improve the Berkeley Function Calling Benchmark by at least 8 points.
- General reasoning: limit GPQA Diamond degradation to 2 percentage points.
Independent reproductions are not cited in the release. Evaluation versions, prompt templates, scoring scripts, sampling settings, context length, quantization, and GPU-offload configuration would all affect a comparison. The reported results should therefore be treated as project measurements until external runs reproduce them.
A 5.7 GB starting point
The quantized repository offers several quality and memory trade-offs. Two practical options are:
| Quantization | Approximate size | Use |
|---|---|---|
Q4_K_M |
5.7 GB | Recommended starting point for limited memory and faster inference |
Q6_K |
7.5 GB | Higher precision with greater memory use |
An 8 GB GPU may hold the Q4 weights, although the key-value cache, context window, runtime buffers, and driver consume additional memory. Long contexts can exceed the available VRAM. Partial CPU offload, a shorter context, or a smaller quantization may be necessary. Apple Silicon uses unified memory, so the operating system and other applications share the same pool.
Current llama.cpp builds can fetch the selected quantization from Hugging Face and start either an API server or an interactive session:
llama-server -hf mradermacher/Cyber-Ornith-1.5-9B-OBLITERATED-i1-GGUF:Q4_K_M
llama-cli -hf mradermacher/Cyber-Ornith-1.5-9B-OBLITERATED-i1-GGUF:Q4_K_M
The server exposes an OpenAI-compatible API at http://localhost:8080/v1 by default. Frameworks such as Hermes Agent and OpenClaw can target that endpoint, but tool schemas, chat templates, stop tokens, and reasoning-tag handling still require integration tests.
Ollama, LM Studio, Jan, and Docker Model Runner can consume GGUF files, subject to each release’s import syntax and feature support. Pinning the exact model file, checksum, runtime version, context size, and prompt template will make deployments easier to reproduce.
Shell access demands hard boundaries
Terminal access turns a mistaken or manipulated response into a system action. Production deployments should place the model behind controls that constrain what its tools can do:
- Run commands inside a disposable container or virtual machine under a non-root account.
- Restrict filesystem access, network egress, credentials, and reachable targets.
- Require human approval for destructive, privileged, or state-changing commands.
- Treat terminal output, web pages, repositories, and scan results as untrusted input that may contain prompt-injection instructions.
- Validate function arguments against strict schemas and explicit allowlists.
- Record tool calls and outcomes while excluding secrets and unnecessary reasoning traces from logs.
- Limit all testing to systems covered by written authorization and a defined scope.
Built for contained red-team work
The training mix combines cybersecurity trajectories with Hermes function-calling v1, OpenThoughts3-1.2M, and synthetic OpenHands conversations. Those sources emphasize structured calls, multistep reasoning, and repeated observe-plan-act loops.
Suitable uses include vulnerability-triage assistants that interpret scanner output, CTF helpers that draft analysis commands, and internal red-team agents operating inside isolated labs. Local execution can keep proprietary code, payloads, and findings away from third-party inference services.
The same specialization creates practical limits. The model can hallucinate flags, package names, vulnerabilities, and command output; visible reasoning does not prove correctness. The authors also describe a trade-off in which CLI and security adaptation reduces some general-purpose breadth relative to the base model.
The model card lists the released weights under Apache 2.0. Developers still need to review the terms attached to base components, datasets, runtimes, and integrated tools before redistribution or commercial deployment.
Local agents move closer to the terminal
Cyber-Ornith combines reduced refusals, security-specific post-training, visible planning, and function calling in a model small enough for consumer hardware. That lowers the hardware and data-sharing barriers for contained security automation while placing command validation and access control in the deployment stack.
Independent evaluation will determine whether the reported benchmark gains survive different prompts, runtimes, and environments. The immediately testable result is a compact local backend that developers can connect to a sandboxed terminal and assess against their own authorized workloads.