AI for security researchers: the questions that get blocked
Penetration testers, bug-bounty hunters, and malware analysts need an AI that can discuss exploits without a disclaimer every other sentence. We tested what the big models refuse.
A penetration tester friend described her workflow last month: open ChatGPT, paste the question, watch the refusal, switch to Stack Overflow, piece together an answer from forum posts older than the CVE she's chasing. This is not a workflow. This is what's left after the workflow was sanitized.
54%
Of security prompts refused by the top consumer chatbot in our test
3%
Refused by Unrestricted on the same test
1 / 200
Security questions that are actually 'how do I hack a specific target'
What actually gets refused
The pattern is bleak. Asking for a walkthrough of how a public CVE works? Refused. Asking for help reading a decompiled binary? Refused. Asking for a canonical example of an SQL injection string for a training lab? Refused. Asking about the mechanics of a published ransomware family's encryption routine? Refused.
These aren't operational attack plans. These are the building blocks of an entire profession. The information is in books, papers, and the SANS curriculum. The model that won't discuss it is the anomaly, not the knowledge.
Security research isn't an offensive discipline. It's the defensive discipline that happens to know what it's defending against.
The floor we keep
There's a narrow band of requests we still won't help with: active exploitation of a named, identified target; live-weaponized malware authored for a specific deployment; social-engineering scripts aimed at a real, named person. Training labs, published research, historical incidents, CVE explainers, decompilation — all fine.
What a useful session looks like
Paste a decompiled function. Ask what it's doing. Ask what the vulnerable pattern would be in clean source. Ask how a fix would look. Ask for an analogy a non-specialist stakeholder would understand, because you also have to report up. No refusals, no moralizing, just the back-and-forth you were going to have with a senior colleague anyway.
Frequently asked
Will Unrestricted help me with a CTF challenge?
Yes. CTFs are training exercises with consenting organizers. We discuss the concepts, the exploitation paths, and the remediation. We won't solve the challenge for you without effort, but we'll teach.
Can I paste malware samples?
Yes for analysis (reading, classifying, reverse engineering). No for 'improve this sample so it evades detection against product X.' The line is offensive authorship against real targets.
What about bug bounty scope?
Fine. We'll discuss methodology, help you reason about findings, draft a report, and weigh severity. We're not a substitute for your own testing, but we're a much faster rubber duck.
Do you keep logs of security-research conversations?
No — we don't store conversations at all, regardless of topic.
Ready to experience an AI without a leash?
Start chatting free