CBRN Specialist, AI Safety Review

What you'll do

Each task puts a real user request in front of you, together with the answer a frontier model gave. You decide whether that answer was safe, accurate and framed the way a professional in the field would frame it, and you write down your reasoning. Sovrano is explicit that the job cuts both ways: a model that hands out operational detail it should have withheld has failed, and so has a model that refuses an emergency planner or a safety officer asking an ordinary question. Expect to spend as much effort on the second kind as on the first.

The panel covers all four hazard classes, but you work in one strand: chemical safety and toxicology, biosafety and biosecurity, radiological protection, or nuclear safety and safeguards. You name your strand in the expert survey Sovrano sends after you apply, so pick the one where you have the most professional hours.

This is Sovrano's misuse and safety framing panel. Specialists in reactor and radiation physics, or in plant and process reasoning, are steered to Sovrano's separate technical panels (Nuclear Physicist and Chemical Engineer).

Who fits

• A bachelor's degree or higher in a CBRN-adjacent field, or equivalent professional experience
• At least 3 years of professional work in one CBRN strand
• Residence in the US, UK, Canada, Australia, New Zealand or an EU member state
• English at C1 or higher, written and spoken
• Time for a paid bootcamp on a Monday, production from Tuesday, and roughly 10 hours of work before the batch closes on Sunday
• Willingness to join Sovrano's Discord before the bootcamp and to invoice Sovrano as an independent contractor

The ad names emergency planners and safety officers as the people models wrongly turn away, and people in those jobs already make disclosure calls at work: what goes into a public notice, a training session or a reply to a worried caller. A reviewer has to see why an answer is unsafe, or why a refusal was unnecessary, without asking you.

What it pays

$35 to $49 an hour, set by your background and confirmed in writing before the bootcamp. Pay is hourly for time worked, and the bootcamp itself is paid. At roughly 10 hours, the first batch comes to roughly $350 to $490 before tax. You invoice Sovrano directly; Sovrano does not file or submit tax forms for you, so tax reporting where you live is your job.

Worth knowing

Good:

• A defined first batch with a start day, an end day and a rate confirmed before you begin (Sovrano's "Expert" waitlist listings attach no project at all)
• The bootcamp is paid
• No fixed shifts or standing meetings; you choose your hours inside the batch window
• Sovrano says there is a good chance the work extends past the first batch, and that it goes back to the people who delivered well

Less good:

• Ten hours is a small amount of certain work, and any extension is a likelihood, not a promise
• The week is tight: a Monday bootcamp, then production that closes on Sunday, so a busy week elsewhere could cost you the seat
• Discord is mandatory, and you are expected to be reachable while the batch runs
• You will read model answers that contain detail they should never have given; if that is uncomfortable, this panel is a poor fit

About this listing

Sovrano posted this role on 21 September 2026. We read it from Sovrano's signed-in job list on 1 November 2026, when it was accepting applications. The rate band, batch structure, residence list and requirements above are the ad's own. The ad gives the weekday pattern of the batch but no calendar dates, so check which week yours falls in when the invitation arrives. See Sovrano AI.

How to get hired

• Apply with a CV or LinkedIn profile, then complete the short expert survey. Only applicants who pass it are offered a bootcamp seat, and the invitation and project guidelines come by email.
• Name one strand in the survey and back it with specifics: the biosafety program you ran, the radiation protection role you held, the chemical safety reviews you signed off. Three years in one strand is the bar, so show the dates.
• Before the survey, take a handful of ordinary questions people ask in your field and write two lines on each: would you answer it, for whom, and how would you frame it. Over-refusal counts against a model as much as over-sharing.
• Clear the batch week in your calendar before you apply, since the bootcamp, production and the Sunday close all fall inside it.
• Our interview prep page covers how AI training screenings usually run.