AI Safety Experts — English & Norwegian

$48–62/hrrate as listed

Red team conversational AI models in English and Norwegian, probing for jailbreaks, prompt injections, bias exploitation, and multi-turn manipulation. Work is text-based, documented against shared taxonomies, and delivered as reproducible attack cases and annotated failure data.

Summary, eligibility guidance and screening preparation are prepared by Sidequest from the platform's posting. Check the source listing before applying.

What the work involves

You run adversarial sessions against conversational models and agents, then write up what broke and why. A typical shift mixes structured probing — working through an assigned taxonomy of misuse categories — with open-ended exploration where you follow a hunch across multiple turns to see whether a refusal can be eroded. Norwegian matters here because safety behavior often degrades outside English: a guardrail that holds in English may fail when the same request is framed in Norwegian, uses Norwegian cultural or legal context, or code-switches mid-conversation.

Outputs are artifacts, not opinions. You annotate failures, classify the vulnerability, rate severity, and produce attack cases another person can reproduce from your notes alone. Some batches cover sensitive material — bias, misinformation, harmful instruction-following. Topics are disclosed before you see content, higher-sensitivity queues are opt-in, and wellness resources are available.

What the platform screens for

Mercor's screen is AI-led and follow-up heavy. Expect it to test whether your red teaming experience is real: naming a technique is not enough, you will be asked to describe an attack you personally ran, what the model did, and why it worked. Native-level Norwegian is verified, and you should expect to be pushed on how safety failures manifest differently across the two languages. The screen also probes judgment — whether you can distinguish a genuine vulnerability from a model being unhelpfully cautious, and whether you write findings a customer engineer could act on.

Logistics

Fully remote, asynchronous, project-based work assigned in batches
Observed pay band of $48–62/hr, varying by project and specialty; not guaranteed
Hours are flexible but batches carry deadlines; most contributors commit 10–20 hours weekly
Assignments rotate across customers, so expect new taxonomies and guidelines periodically

What they look for

Native fluency in both English and Norwegian
Prior red teaming experience — AI adversarial testing, penetration testing, or socio-technical probing
Ability to document attacks reproducibly, with clear severity and vulnerability classification
Working knowledge of LLM failure modes: jailbreaks, prompt injection, multi-turn manipulation
Comfort applying structured taxonomies and benchmarks rather than ad hoc testing
Willingness to review sensitive content covering bias, misinformation, and harmful behavior

Before you apply

Would you pass this screen?

Drawn from what this listing actually requires. Answered in your browser — nothing is sent to us, and we do not store your answers.

Are you a native English speaker, or do you have native-level fluency in English acquired in childhood?
Native fluency in English · required
Are you a native Norwegian speaker, or do you have native-level fluency in Norwegian acquired in childhood?
Native fluency in Norwegian · required
Have you completed at least one professional project or role that involved red teaming an AI model, testing an AI system for adversarial vulnerabilities, or generating adversarial datasets?
Red teaming or adversarial AI experience · required
Do you have prior professional experience in cybersecurity, penetration testing, or structured adversarial testing of any system (not necessarily AI)?
Security or adversarial testing background · preferred

0 of 4 answered

Included with this listing

Interview prep — 11 questions

Matched to what this platform's screening actually measures. Answer each out loud in 60–90 seconds: claim → concrete example → implication.

What it tests: Whether the candidate has genuine hands-on red teaming or adversarial testing experience, and whether they can articulate a specific vulnerability discovery process rather than speaking in abstractions.
A strong answer: A concrete example naming the system (AI model, application, or security surface), the specific attack vector they devised (jailbreak technique, prompt injection, social engineering angle, etc.), and what the vulnerability revealed—not a generic 'I found a bug' but evidence of methodical probing.
Where people lose it: Describing theoretical knowledge of attacks without an actual case where they implemented one, or citing only participation in a bug bounty program without explaining their own creative input in finding the flaw.