Remote AI Safety Red Teamer

Mercor; Lisboa; À distance; EUR 90 000 - 130 000; Plein temps

Résumé du poste

Mercor is seeking experienced AI Safety Red Teamers to identify vulnerabilities in frontier AI systems through adversarial testing. You will design challenging prompts, uncover model weaknesses, and evaluate AI behavior across high-risk and ambiguous topics.

Responsibilities include designing adversarial prompts, identifying jailbreaks, unsafe behaviours, and policy failures, and documenting vulnerabilities to contribute to safety benchmarking.

Qualifications

• Bachelor's degree or higher in a related discipline or equivalent.
• 5+ years of professional experience in AI Safety, AI Red Teaming, Trust & Safety, cybersecurity, investigative journalism, life sciences, or related field.
• Strong analytical reasoning, prompt design, and written communication skills.
• Experience designing adversarial prompts or evaluating frontier AI systems.

Responsabilités

• Design adversarial prompts to stress-test frontier AI models.
• Identify jailbreaks, unsafe behaviours, hallucinations, and policy failures.
• Evaluate model robustness across misinformation, cyber, biosecurity, fraud, political content, and other sensitive domains.
• Document vulnerabilities and contribute to safety benchmarking and red-teaming reports.
• Collaborate with AI researchers to improve model alignment, robustness, and safety.

Connaissances

Adversarial testing

Prompt design

Model evaluation

Analytical reasoning

Formation

Bachelor's degree or higher

Outils

RLHF

Jailbreak testing

Description du poste

Mercor is seeking experienced AI Safety Red Teamers to identify vulnerabilities in frontier AI systems through adversarial testing. You will design challenging prompts, uncover model weaknesses, and evaluate AI behavior across high-risk and ambiguous topics.

Responsibilities include designing adversarial prompts, identifying jailbreaks, unsafe behaviours, and policy failures, and documenting vulnerabilities to contribute to safety benchmarking.