Job Description

Refusal Preference Reward Model Evaluator is a remote red-team track for stress-testing AI systems against adversarial prompts. Reviewers craft attack scenarios, document the failure mode, and pair each successful jailbreak with the rubric clause it violated so the safety team can patch the gap.

Why this role matters

Adversarial evaluation is how AuraOne hardens AI models before they ship to customers. Reviewers think like attackers and write up failures with enough rigor that the modeling team can reproduce, fix, and regress-test them.

Responsibilities

• Design adversarial prompts that probe known weakness classes (jailbreak, policy bypass, prompt injection) for Refusal Preference Reward Model Evaluator assignments.
• Document every successful attack with reproduction steps and the policy clause it violated.
• Score model defenses across single-turn and multi-turn conversations.
• Triage emerging attack vectors and route them to the safety team with severity ratings.
• Maintain a personal library of attack patterns and propose new red-team rubrics.
• Calibrate against the broader red-team cohort to keep coverage and severity consistent.

Qualifications

• Demonstrated experience red-teaming AI systems, security research, or adversarial ML work for Refusal Preference Reward Model Evaluator work.
• Strong written communication — your reports become the patch ticket.
• Comfort working in policy-grey areas with clear documentation of what was attempted and why.
• Familiarity with prompt-injection, jailbreak, and policy-bypass taxonomies.
• Reliable async availability for at least 10 hours per week.

Example tasks

• Construct a 5-turn adversarial conversation that bypasses a specific policy clause and write up the patch ticket.
• Score a model's defenses against a known jailbreak pattern across 20 variants.
• Propose a new red-team rubric category after spotting an emerging attack vector.
• Reproduce a failure another reviewer reported and confirm the severity tag.

Nice to have

• Background in offensive security, AppSec, or trust & safety operations.
• Experience publishing or reproducing public adversarial-ML research.
• Multilingual fluency for cross-language attack testing.

Skills

• Adversarial prompting
• Red-team analysis
• Policy taxonomy
• Failure documentation
• Refusal Preference Reward Model evaluation
• Preference ranking
• RLHF
• Rater calibration
• Refusal
• Preference

Work model

Remote — US-eligible. Remote · Independent specialist contractor. Employment type: CONTRACTOR. Applicants must be authorized to work from US.

Compensation

Hourly rate confirmed after the interview process.

Application process

Apply through AuraOne's specialist intake for role-specific routing and review. Final project scope, schedule, and contractor terms are confirmed before placement.

Job Tags

Similar Jobs