AI Red Teamer, Cybersecurity (Remote)

Cybersecurity Red Teamer

Location: Seattle, WA, or Remote within the United States

Work arrangement: Flexible, including part-time availability

Handshake was founded on a simple belief: everyone deserves a path to a great career, regardless of where they went to school or who they know. Today, we support 25 million job seekers, more than 1 million employers, and 1,600 educational institutions.

In 2025, we started Handshake AI and built the fastest-growing AI data business in history. We work directly with researchers at frontier AI labs to create evaluations, publish benchmarks, and push the boundaries of data. We have grown from $0 to approximately $1 billion in run rate and pay approximately $60 million to more than 30,000 individuals every month.

Why join Handshake now

• Shape how careers evolve in the AI economy at a global scale, with an impact your friends, family, and peers can see and feel
• Partner closely with world-class AI labs, Fortune 500 companies, and leading educational institutions
• Work with engineers, scientists, operators, and other professionals from organizations such as Palantir, Meta, and Scale AI, as well as former YC founders
• Help build a massive, rapidly growing business with billions in revenue
• Work from Seattle or remotely from anywhere within the United States

About Handshake AI Human data is core infrastructure for AI advancement. Frontier AI labs currently improve model capabilities through data-intensive post-training techniques. We believe spending on AI training data will increase three to five times over the next few years and continue growing as models expand into new domains.

Handshake AI supports frontier AI labs by working on their most complex data challenges at scale.

About The Role As a Cybersecurity Red Teamer, you will evaluate whether AI models can be manipulated into generating functional malware, viable exploit code, attack tooling, or step-by-step operational guidance that could give a threat actor meaningful assistance in carrying out cyberattacks.

Your job is to find the gaps between what a model’s safety guardrails are intended to block and what a skilled adversary can actually extract.

This role requires you to think like an attacker who has access to a highly capable AI assistant. You will craft adversarial prompts and multi-turn interaction chains that simulate how real threat actors, ranging from inexperienced attackers to advanced persistent threat operators, might use LLMs to accelerate reconnaissance, weaponization, exploitation, lateral movement, persistence, and exfiltration.

You will then evaluate whether the model’s output is genuinely dangerous or merely surface-level noise.

Deep cybersecurity expertise is essential. Your value will come from being able to examine a model-generated payload, exploit chain, or attack plan and determine whether it would actually work, how much refinement it would require, and what type of attacker it could meaningfully assist.

This position may be performed from our Seattle location or remotely within the United States. Seattle-based and remote team members will collaborate closely through shared evaluation workflows, regular feedback, and virtual working sessions.

What you will do

• Design technically grounded adversarial prompts that test whether models provide meaningful assistance across the cyber kill chain, from reconnaissance through exfiltration and impact
• Evaluate model-generated code and technical output for functional correctness, determining whether outputs represent real exploits, plausible attack tooling, or nonfunctional noise
• Test model behavior across offensive categories, including malware generation, vulnerability exploitation, social engineering, credential harvesting, privilege escalation, command-and-control infrastructure, and data exfiltration
• Probe dual-use boundaries by testing how models respond to queries that combine legitimate security research, penetration testing, and defensive operations with offensive applications
• Simulate attacker personas at varying skill levels, including opportunistic, intermediate, and advanced or APT-level actors
• Test multi-step and multi-turn attack chains, including scenarios in which early turns establish benign context before pivoting to malicious requests
• Score model responses using structured harm taxonomies and severity rubrics calib…