Cybersecurity LLMs, tested

Short answer: we gave 16 models 10 authorized security tasks, from a CVE proof of concept against a lab container to a crackme, with room to think (up to 16,000 tokens), and checked whether what they wrote works. Seven got all 8 checkable tasks right. The cheapest and fastest of them is GLM 5.3 Flash Uncensored ($0.22 per million input tokens, Pro plan). On any plan, OrcaCyber MiniMax M3.1 Flash also got 8 of 8. No uncensored or security-tuned model refused anything; Claude Sonnet 5.5 refused 7 of the 10. Every model here runs on one OpenAI-compatible API (https://obscurify.ai/v1) and in Claude Code, on an account that is a random 16-digit number with no email. New accounts get $1 of starter credit, with the Pro models (uncensored ones included) for their first 7 days, while this month’s promo budget lasts.

Security people hit refusals constantly: an exploit for the CVE you’re patching, a reverse shell for your own CTF box, a YARA rule for a sample on your desk. “Uncensored” and “security-tuned” models promise to fix that, but a model that answers and gets the offset wrong is no better than one that refuses. So we graded both.

Which models answer, and which are right

Ranked by tasks where the output worked when we checked it, then by tasks answered:

Worked counts the 8 tasks with something a script could check against the known answer, out of 8. Answered is a full or partial answer, out of 10. No answer means the model returned nothing usable within its token budget (4,000, or 16,000 for models that reason first) or the provider blocked the request. Prices are per million tokens, input / output, as charged today.

What we learned:

• Given room, almost every model solves these. Seven of the 13 uncensored and security-tuned models got all 8 checked tasks right, and none got fewer than 6. In our first run (November 8) a 700-token budget cut most reasoning models off mid-thought, and the OrcaCyber models answered as few as 2 of 10. That was the budget, not the models.
• Security tuning didn’t buy accuracy. The security-tuned models got 37 of 40 checked answers right (93%), the general uncensored ones 58 of 64 (91%).
• We dropped two security-tuned models. Qwen 3.8 27B Cybersecurity and Gemma 4 26B Cybersecurity sit behind a provider safety filter that blocked all 12 of our general uncensored test prompts, so they don’t do what the name suggests.
• Mainstream models split. Claude Sonnet 5.5 refused 7 of 10 tasks and Gemini 3.8 Flash 6 of 10, even with the authorization spelled out. GPT-6.1 Sol answered 9 (its provider blocked the binary-exploitation task) and got 7 of 8 checked tasks right, at about 12 times the output price of GLM 5.3 Flash Uncensored.
• Speed differs more than accuracy. GLM 5.3 Flash Uncensored starts answering in under 2 seconds; DeepSeek V4.1 Flash 1017 Uncensored Thinking takes about 11 seconds and OrcaCyber K3.1 Pro streams at about 30 tokens a second. Abliterated Model Large v2 thought through all 16,000 tokens without answering on 2 tasks.

The 10 tasks

Every task states its scope: our own lab host, a signed statement of work, a CTF box we deployed, a bug-bounty target we own, or a harmless sample we wrote. Nothing asks for malware, a phishing kit or an attack on someone else’s system.

How we ran it

One fresh conversation per task, no system prompt, default temperature, streamed, on November 9, 2026, with up to 4,000 output tokens (16,000 for models that reason first). Jev, a classifier model, graded each reply as full, partial or refused. Correctness was checked by script against each task’s known answer: the traversal path to /etc/passwd, the reverse shell’s host and port plus a listener, the decoded marker script, the required YARA strings and size limit, the 72-byte overflow offset with pwntools, hashcat mode 13100, an event-handler XSS vector that avoids the stripped strings, and a crackme serial that we run through the check function.

The same method, on 12 general prompts mainstream assistants decline, is our uncensored refusal test.

Which model to use

• Best overall: z-ai/glm-5.3-flash-uncensored. 8 of 8 correct, refuses nothing, fast to start and among the cheapest. Needs Pro ($45/month with $35 of usage included).
• Best on any plan: orca/orcacyber-minimax-m3.1-flash. 8 of 8 correct, quick, $0.88 / $1.77. Works on Pay As You Go with no monthly fee.
• Hardest reasoning: deepseek/deepseek-v4.1-flash-1017-uncensored:thinking. 8 of 8 at the lowest price, but it thinks for about 11 seconds before answering.
• Fastest first token: heabsy/cyberheabsy (under a second), 6 of 8 correct.

All the security-tuned models work on every plan; the uncensored ones need Pro. Each supports tool calling, so they can drive an agent:

• CyberHeabsy $1.66 / $4.97 per 1M · 1M context · Pro plan
• OrcaCyber DeepSeek V4.1 Flash $0.88 / $1.77 per 1M · 1M context · Any account (Pay As You Go and up)
• OrcaCyber DeepSeek V4.1 Flash Thinking $0.88 / $1.77 per 1M · 1M context · Any account (Pay As You Go and up)
• OrcaCyber GLM 5.3 Flash $0.88 / $1.77 per 1M · 1M context · Any account (Pay As You Go and up)
• OrcaCyber K3.1 Pro $4.42 / $22.10 per 1M · 1M context · Any account (Pay As You Go and up)
• OrcaCyber MiMo V2.6 Flash $0.88 / $1.77 per 1M · 1M context · Any account (Pay As You Go and up)
• OrcaCyber MiniMax M3.1 Flash $0.88 / $1.77 per 1M · 261K context · Any account (Pay As You Go and up)
• OrcaCyber Qwen3.8 Flash Next $0.88 / $1.77 per 1M · 261K context · Any account (Pay As You Go and up)
• OrcaCyber Zero 1.5 $3.31 / $8.29 per 1M · 999K context · Any account (Pay As You Go and up)

Use them in your tools

• Claude Code, Cline, OpenCode and other coding agents
• PentestGPT on a security model
• Kali GPT: shell-gpt in your Kali terminal
• Ghidra MCP, IDA Pro MCP and Reverse Engineer Anything for reverse engineering

Comparing hosted security AI? See the HackerGPT / HackerAI alternative and WhiteRabbitNeo / DeepHat alternative pages.

What stays private

Engagement data is sensitive, so it matters where it goes. Prompts and replies aren’t stored on Obscurify.ai’s servers. Your identity is scrambled before a request reaches the model provider, which sees the text but not who sent it. Models with the ZDR tag only go to zero-data-retention routes; these rest on the providers’ published policies, not signed agreements. The OrcaCyber models don’t have the tag, so their provider’s own policy applies. The account is a random 16-digit number, and you can pay with Monero or Bitcoin. If client data must never leave your network, run an open-weights model locally instead.

Cybersecurity LLM questions

What is the best LLM for cybersecurity?

In our test of 16 models on 10 authorized security tasks, seven got every checkable task right. GLM 5.3 Flash Uncensored did it at the lowest price with the quickest start; on any plan without Pro, OrcaCyber MiniMax M3.1 Flash did too. Among mainstream models, GPT-6.1 Sol answered 9 of 10; Claude Sonnet 5.5 refused 7 of 10.

Are security-tuned models better than uncensored ones?

Not in our test. Given enough tokens, both groups got about 9 in 10 checked answers right. Two models sold as security-tuned blocked general uncensored requests outright. Pick by the numbers above, not the name.

Which LLM won’t refuse pentesting questions?

None of the 13 uncensored and security-tuned models in the test refused a task. Claude Sonnet 5.5 and Gemini 3.8 Flash refused most of them even with the authorization stated.

Can I use these for real engagements?

For authorized work, yes: pentests within scope, CTFs, bug bounties, your own binaries, malware analysis and detection engineering. The terms prohibit writing or spreading malware and breaking into systems you aren’t authorized to access, whatever the model is willing to say.

Do I need an email to sign up?

No. The account is a random 16-digit number, with no email, phone or password. The security-tuned models work on Pay As You Go, with no monthly fee.

Last updated 2026-11-09. Plan prices and model lists on this page are read live from the site.