Senior AI Product Manager, Cybersecurity

Scale has been the leading AI data foundry, helping fuel the most exciting advancements in AI, including frontier model training, enterprise adoption, defense applications, and autonomous vehicles. Our mission is to develop reliable AI systems for the world's most important decisions We're looking for a Senior AI Product Manager to build and own Scale's Cybersecurity portfolio — the data, environments, and evaluations frontier labs use to train and measure security capability in their models. This is a build role: you will define the strategy and standards for a product line that does not exist yet. Security is where the hardest problems in agentic AI now sit. An agent that can find a vulnerability, prove it reproduces, and patch it without breaking the system is doing work that takes a skilled human days. Measuring that honestly requires reproducible execution environments at scale and practitioners who have actually done the work. Scale has the first, proven across SWE-Bench Pro, SWE Atlas, and our contributions to the Terminal-Bench lineage. You will build the second. You Will Own the roadmap and strategy for Scale's Cybersecurity portfolio across training data, RL environments, agentic task suites, and evaluation products — and stand the product line up end to end, from task taxonomy and sourcing through pricing and first external release. Define the capability map we train and measure against: vulnerability discovery, proof-of-concept reproduction, patch generation and regression safety, secure code review, supply-chain analysis, malware and binary analysis, detection engineering, and incident triage. Make the strategic call on where Scale competes across the offense–defense spectrum — which capabilities we build training data for, which we only measure, and which we decline. Partner with ML researchers and security practitioners on task specifications, grader design, and verifiable rewards, holding to execution-grounded verification wherever possible: a task counts as solved only when the reproducer fires or the patch holds without breaking functionality. Drive the infrastructure roadmap — reproducible vulnerability images, fuzzing and build toolchains, sandboxed execution, network-segmented ranges, automated verification — and build sourcing pipelines that scale past hand-curation. Own the responsible-development posture: containment, coordinated disclosure for live vulnerabilities surfaced during task construction, need-to-know handling of sensitive artifacts, and customer vetting, working with Security, Legal, and Policy to make these processes real rather than nominal. Establish governance for data quality, contamination prevention, license and IP hygiene, reproducibility, and release management. Recruit and steward a contributor network of working practitioners — vulnerability researchers, exploit developers, malware analysts, detection engineers, incident responders — and design quality controls that hold up when reviewers are validating work at the edge of their own expertise. Own external partnerships across open-source benchmark collaborations, academic security groups, and enterprise data partnerships. Work directly with frontier labs and enterprise customers to understand where their models fail on security work, translate that into roadmap, and partner with GTM on launches and thought leadership. Ideally, You'd Have Real cybersecurity work under your belt, rather than security-adjacent product experience: vulnerability research, fuzzing and crash triage, reproducer development, patch and root-cause analysis, exploit development, malware analysis, red teaming, detection engineering, or incident response. Competitive CTF, published CVEs, a bug bounty record, or OSS-Fuzz contributions all count. 5+ years in product management, technical program management, consulting, or customer-facing technical roles — or equivalent depth as a practitioner with a clear pull toward product ownership. A working view of the AI-for-security evaluation landscape and where it falls short: CyberGym, Cybench, CVE-Bench, BountyBench, CyberSecEval. CyberGym sets the bar we hold ourselves to — real vulnerabilities sourced at scale, execution-grounded verification, and tasks hard enough that frontier agents still clear only about a fifth of them. Enough software engineering depth to read unfamiliar code, reason about runtime behavior, and hold your own with senior engineers and ML researchers. Familiarity with how models are post-trained and evaluated, including agentic scaffolds and container-based rollout infrastructure. Excellent stakeholder management and executive communication skills, with a demonstrated ability to drive alignment across cross-functional organizations. Sound judgment on dual-use questions, and genuine care about building capability measurement that helps defenders more than attackers. Entrepreneurial mindset, bias for action, and comfort operating