It started with a simple question about an open-source security repo.
A few prompts later the conversation had drifted into attack-surface mapping, testing logic, wireless lab setups, and the general shape of how somebody with enough patience could assemble a workflow they did not invent themselves.
That is the part worth paying attention to.
Not because I want to run crime fan-fiction through a chatbot. Because if you are serious about recovering stolen crypto, tracing scams, or building a small company around post-incident response, you need to understand how quickly modern assistants can help people organize bad intent into something that feels operational.
Same machine. Two uses. Build faster, or get worse faster.
Ten Ordinary-Looking Prompts Into Attacker-Shaped Territory
The useful insight was not any one answer. It was the progression.
The conversation moved like this:
- Ask what a red-team repo is.
- Ask how you would test your own site.
- Ask what an Alfa adapter teaches you.
- Watch the assistant start laying out tooling, sequences, lab habits, and attacker-adjacent thinking in a calm helpful voice.
The model did not need a direct prompt that said, "teach me how to steal."
It responded to something softer:
- curiosity
- self-testing language
- lab framing
- "I own this" framing
- step-by-step escalation
That is how a lot of real misuse happens now. Not with one cartoonishly evil prompt. With ten ordinary-looking prompts in a row.
Each Answer Handed the Next Category to the Attacker
The questions were framed as legitimate. "What is this tool?" "How would I test my own site?" "What can I learn with this hardware?" Those all sound ordinary. In many cases they are ordinary. A security researcher, a sysadmin, a founder, and a bored criminal can all ask the same question. The model has to answer the surface intent first.
Each answer became the next scaffold. You do not need one perfect prompt if each answer hands you the next category: recon, scanning, auth testing, injection, lab hardware, packet capture, protocol awareness. The answer itself becomes the outline for the next round.
Tool names are retrieval anchors. Once a conversation picks up names of common tools, frameworks, and workflows, the assistant has more structure to pull from. That does not mean the operator suddenly became a real expert. It means the model started handing them a shape. For a bad operator, shape is often enough.
The tone stays neutral while the implications do not. An assistant can describe ugly things in a clean, professional, almost educational tone. That tone makes the material feel safer and more legitimate than it really is. That is one reason red teaming the model matters. The danger is not only what it says. It is how calm it sounds while saying it.
The Theft Is Rarely Just "the Blockchain Part"
Because crypto theft is rarely just "the blockchain part."
People lose money through:
- wallet drain approvals
- seed phrase theft
- fake support flows
- impersonation
- malicious contract interactions
- social engineering
- exchange off-ramping
- timing, laundering, and chain-hopping after the initial hit
If you want to help victims, you need to see the attacker stack for what it is:
not magic
not genius
not always deeply technical
Often it is just workflow.
A phishing page here. A fake urgency cue there. A bad signature request. A few hops. A cash-out path. Enough pressure and confusion to keep the victim from reacting in time.
That is why red teaming the assistant is useful. The question is not "how do I attack?" It is: what kind of sequence can an attacker assemble quickly now, and what does the defense side have to be ready for?
The Lawful Version
If you were building a crypto recovery company starting today, the real work is not glamorous.
It looks more like this:
