Tech • AI • Robotics • Game

ENFR

I tested local AI with no limits: Offensive cybersecurity - (Complete tutorial)

7/10
AIParlons IANovember 20, 2026 at 02:13 PM28:35
Audio player

⏱ ≈ 3 min read · 29-min video — 26 min saved

TL;DR

Uncensored and weakly restricted AI models can help automate parts of cybersecurity work, but they also lower the barrier to finding exposed secrets and testing attacks, raising legal and security risks for poorly built applications.

KEY POINTS

Low-friction attack testing

Testing with several models showed sharply different guardrails. Kimi 3.1 Pro refused a request to run API-key retrieval, while Minimax M3 Flash proceeded and surfaced exposed credentials from public code searches. In one demonstration, scanning 150 application repositories produced 10 extracted keys, underscoring how easily secrets can leak when developers publish unsafe code or configuration files.

Uncensored local models are accessible

So-called Heretic variants of Qwen 2.5 and Qwen 3 were presented as low-refusal models that can run locally through Open WebUI. Hardware determines model size: a 27B model was said to require about 32 GB of VRAM, roughly the class of an Nvidia RTX 4090, while smaller 3.8B or 7B versions can run on GPUs with around 10 GB. The Qwen 2.5 32B Heretic Obliterated build was described as one of the least restricted options.

Refusal rates vary widely

Mainstream models were described as heavily trained to reject unauthorized or dangerous requests, while uncensored builds were framed as having only 1% to 22% blocking, versus near-total refusal for standard safety-tuned systems. That gap matters because the same prompt can be rejected by one model and executed by another, making model selection a core operational and governance issue.

Local AI is not truly unlimited

Despite claims that local AI can replace frontier systems, practical limits remain steep. Even expensive hardware was described as offering only about 256K context, far below the 1 million-token context windows available from some remote services. That makes local models useful for smaller tasks and flexible prompting, but less suitable for full-scale audits of large applications or websites.

Cyber audits require multiple models

Security work was framed as a multi-model pipeline rather than a single-chatbot task. Benchmarks cited included CyberGym, SecBench Pro, ExploitBench, ExploitGym, CWE, and Cyber Defense Benchmark. Different models reportedly perform best on different tasks: DeepSeek variants for some security benchmarks, Mimo V2 Flash for CyberGym, and Kimi for exploit-oriented testing.

Detection and repair are separate problems

A central warning was that a model good at finding flaws may be weaker at fixing them. Effective auditing therefore requires detection, patch proposal, redeployment, and retesting. That process must also confirm that a fix does not break core functions, which is why backup, regression testing, and repeated security validation are necessary.

Agentic automation changes the economics

Rather than running checks manually, an AI agent orchestration setup was described as capable of performing roughly 400 tests over several hours or days. That approach can compress work previously done by teams, but it also raises the bar for technical skill. The implication is that productivity gains will favor operators who can build workflows, memories, and recursive multi-pass loops rather than those relying on one-shot prompts.

Compliance remains a major barrier

The broader argument was that generating code or websites quickly is far easier than shipping a compliant product. In Europe, especially under GDPR and related data-protection rules, organizations are expected to provide security checks, architecture planning, and ongoing maintenance. A chatbot or AI-generated site can become a liability if it exposes data, lacks audit trails, or cannot withstand legal review after a breach.

Cheap demos can mask costly weaknesses

The examples highlighted a recurring problem in AI-driven software development: polished front ends can conceal insecure architecture. Public-facing apps built through rapid prompt-based methods may appear impressive but still expose databases, secrets, or unsafe code paths. In that context, “vibe coding” was portrayed less as a shortcut and more as a potential route to fines, breaches, and operational failure.

CONCLUSION

Advanced AI can accelerate both cyberdefense and offensive testing, but the same tools also make insecure development easier to exploit. The decisive factor is no longer access to AI alone, but the ability to audit, patch, and govern systems to a professional standard.

Ask a question

More from AI