AI Safety Experts — English & Norwegian

$48–62/hr · Mercor · Hourly, 40 hours a week

Conduct AI red teaming in English and Norwegian to identify vulnerabilities and generate safety training data.

What you would do

Design adversarial scenarios targeting model biases, misuse potential, and manipulation vulnerabilities
Apply jailbreak and prompt injection techniques systematically following documented playbooks
Create multi-turn conversations that probe model consistency and value alignment
Generate annotated datasets classifying failures by vulnerability type and severity
Document reproducible attacks and their implications for real-world deployment

Prior red teaming, adversarial ML, or cybersecurity experience
Native fluency in both English and Norwegian
Curiosity and systematic approach to finding model breaking points
Skill at articulating vulnerabilities and security implications to both technical experts and general audiences
Adaptability across different model types and customer contexts

Red teamingAdversarial input designJailbreak techniques

1.Novel attack generation
Existing playbooks become stale; your ability to innovate determines long-term data value.
For example: “A model has been hardened against direct jailbreaks. Describe a multi-turn conversation approach that exploits different vulnerabilities in sequence.”
2.Cross-lingual bias detection
Biases manifest differently in Norwegian versus English; bilingual expertise matters.
For example: “Design a scenario to test whether a model exhibits different gender stereotypes when responding in Norwegian versus English.”
3.Distinguishing benign from harmful ambiguity
Not every surprising output is a vulnerability; your judgment guides training data quality.
For example: “A model gives an odd response to a Norwegian idiom. How would you assess whether this is a concerning gap or expected behavior?”
4.Documenting reproducibility
Training data must contain actionable, reproducible findings.
For example: “You found an attack but it works inconsistently. How would you modify your approach to make it reliable and document it?”

Conduct red teaming on a provided prompt: find 3-4 distinct vulnerabilities, document each with reproducible steps and risk assessment.

Study existing red teaming frameworks and taxonomies for classifying failures
Review examples of jailbreaks, prompt injections, and bias exploitations
Think about cultural or linguistic differences that might create vulnerabilities in Norwegian vs English
Prepare 2-3 original attack ideas you could articulate step-by-step

We wrote this page from the public Mercor listing. It may be out of date, so read the full posting before you apply.