The work, field by field.

Exploit analysis, malware reverse engineering, and phishing simulations for authorised engagements, without arguing that you are the good guy.

The refusal lands on the word, not the engagement. A scoped test still needs the bug named in enough detail that the owner can fix it. Stock models sand the finding down, or they stop and ask for proof of virtue. The report comes back too vague to reproduce.

Here the model reads the trace and drafts the finding. The authorisation does not move with the weights. The system has to be yours, or the owner has to have allowed the test. Point it at anything else and the account is closed.

Toxicology, reaction hazards, controlled-substance pharmacology. The routine questions that trip keyword filters long before they reach a real hazard.

A toxicology review and a street recipe share a vocabulary, so the filter blocks the compound name. Inside a lab the questions are ordinary. What this intermediate does to the liver. Where this reaction runs away. Which impurity the regulator is going to ask about.

The model stays on the question in front of it. It is not a bench guide for making a poison or a controlled substance. Using the service to harm someone closes the account.

Threat modelling, weapons-effects analysis, wargaming, and adversary OSINT. Topics a consumer chatbot refuses on sight.

Those documents are the staff job. A consumer model refuses the estimate, the wargame turn, and the read of published doctrine before it has looked at the question. The work sits there anyway, and someone has to write it.

The model will draft the estimate and argue the other side of the plan. It will not help you pick a person and go after them.

Crime fiction, war reporting, horror, and true crime. Dark material written as dark as the story needs, with no disclaimers in the prose.

Stock models soften the violence, bolt a warning onto the paragraph, or refuse the chapter. The page comes back looking like a content policy. A novelist or a reporter cannot ship that.

The draft here is the one you asked for. The ban on sexual content involving minors does not move because the request is a chapter. Neither does the ban on using the model to target a real person.

Persuasive copy, competitor teardowns, and provocative campaigns, minus the lecture about manipulation.

Persuasive is the job. A stock model hears the word and writes a lecture, then a paragraph so balanced it cannot run. Competitor teardowns get the same sanding. The sharp line comes off because sharp might be unfair.

The model here writes the line you would actually buy media for. It does not write a scam, and it does not impersonate a real person to take their money.

Frank, accurate information on drug use and overdose, written for people who will use it anyway. No moralising.

Outreach workers and clinicians need plain language on what a substance does, what an overdose looks like, and what to do in the room. Stock models refuse, or they paste a warning and end the turn. The person in front of you leaves with nothing they can use.

The model answers the factual question. It will not help someone make the drug, sell it, or press it on a person who did not ask.

Analyse extremist propaganda, fraud schemes, and leaked documents. Read the worst of the internet so you can report on it.

The story is often the file a model will not open. A manifesto, a scam page, a leak. Stock models refuse the document or return a summary too clean to report from. You cannot check a claim you were not allowed to read.

The model reads the file and pulls the claims that do not hold. The byline stays yours. Reporting the document is the work. Going after a private person with it is the ban.

For clinicians and support services to discuss self-harm, eating disorders, and addiction candidly, instead of the model pasting a hotline and ending the conversation.

Inside a supervised session that paste is the failure. A counsellor or a clinician has to stay with the words the person used, and then write the note. Stock models treat the turn as a crisis they must shut down. The session ends. The note does not get written.

The model stays on the clinical question. It will not provide a method. It is not a substitute for a person when someone is in danger.

Reconstruct how an offence was alleged, test the prosecution's timeline, and read graphic disclosure without the model redacting the file.

From outside the prompt, defence work looks like the crime. Stock models refuse to help you "plan" a thing you are in fact defending, or they black out the pages the case turns on. The brief stalls where it matters.

The model works those pages. It will not help someone commit the offence. An account that asks for that is closed.

Label hate speech, scams, and graphic content at scale. A moderation model has to read what it is filtering.

A classifier that flinches cannot moderate. The queue is the content, and someone has to mark it. Stock models refuse the item they are supposed to sort, or they describe it so vaguely the label is a guess.

The model reads the item and returns the call. Writing the abuse yourself and sending it at real people is the thing the policy is there to stop.

Generate adversarial prompts and attack data to stress-test your own models, classifiers, and guardrails.

A guardrail tested only on polite prompts is untested. The eval is the attacks, the cases your classifier should catch, the set a stock model refuses to write. You end up benchmarking questions nobody was going to ask.

The model will write that set for a system you own. Breaking into someone else's product with it is the same ban as any other unauthorised attack.

Map laundering typologies and synthetic identities so compliance teams recognise them first.

The typology is the control. How a laundering pattern shows up in payments. How a synthetic identity gets through a check. Compliance needs that written down before a customer does it. Stock models treat the request as the crime and refuse the training material.

The model writes the pattern and the red flags. It will not write a script to aim at a victim.