AI Safety Experts: English & Danish

What you'll do

Mercor is building a red team for a client project and wants Danish speakers in it. The job is to attack conversational AI on purpose: try to talk a model past its safety rules, slip instructions into content it processes (prompt injection), steer it towards biased or misleading output over a long conversation, and find the misuse cases its developers did not anticipate.

Every successful break becomes a record. You annotate what failed, classify it against the project taxonomy, note whether it looks like an isolated slip or a pattern, and write it up so the client's engineers can reproduce it. The ad stresses playbooks and benchmarks, so this is closer to a test campaign than to freelance hacking.

Models are usually tuned hardest on English, and a refusal that holds in English can fold when the request arrives in Danish, in mixed Danish and English, or wrapped in local cultural references. Recognising when a Danish answer crosses a line (or when a model's idea of Danish politics or society is subtly skewed) needs a native reader.

The work is text-only. Sensitive topics such as bias, misinformation and harmful behaviour are part of it; you are told the topics before seeing any content, and the higher-sensitivity projects are optional with guidelines and wellness support.

Who qualifies

The ad asks for:

• Native fluency in Danish and English
• Earlier red-teaming work, broadly defined: AI adversarial testing, cybersecurity, or socio-technical probing
• An instinct for pushing systems until they break, combined with a disciplined, framework-based approach
• The ability to explain a risk clearly to both engineers and non-specialists
• Adaptability across projects and clients

Pluses: adversarial ML (jailbreak datasets, prompt injection, RLHF/DPO attacks, model extraction), pentesting and exploit or reverse-engineering experience, harassment and disinformation analysis, and unconventional backgrounds in psychology, acting or writing.

There is no stated degree or years-of-experience minimum, but "prior red teaming experience" is listed as who you are, not as a nice-to-have. If your background is language work only, this is not the easiest entry point.

What it pays

$48–62 per hour, as published by Mercor. Danish sits in the same band as the Norwegian, Swedish, Finnish and Dutch versions of this listing, the highest in the family.

The ad states weekly payment through Stripe or Wise, based on services rendered, and says Mercor cannot support H-1B or STEM OPT candidates.

Worth knowing

Good:

• Among the best-paid language variants of this role
• 142 hires on this listing this month, so the project is clearly staffing up
• No country restriction is published; any Danish and English native speaker can apply
• Topics disclosed in advance, and the most sensitive work is opt-in
• Own schedule, fully remote

Less good:

• The experience bar is real: expect screening to test adversarial thinking, not just language
• Exposure to disturbing text is inherent to red-teaming, however well supported
• Hours and project length are not stated, and projects can end early
• Contractor status means no benefits or employment protections
• Clients and target models are not disclosed

About this listing

Posted by Mercor as a remote hourly contract and read from the Mercor site on 24 September 2026. Pay, requirements, payment terms and the visa limitation come from the ad itself. Mercor notes the work does not involve confidential information from any employer, client or institution. See Mercor.