Technical Product Manager, Trust & Safety

Innodata (Nasdaq: INOD) is a global data engineering company. We believe that data and Artificial Intelligence (AI) are inextricably linked. Our mission is to enable the responsible advancement of artificial intelligence by providing the data, evaluation frameworks, and human expertise required to build AI systems that can be trusted at scale. We provide a range of transferable solutions, platforms, and services for Generative AI / AI builders and adopters. In every relationship, we honor our 36+ year legacy delivering the highest quality data and outstanding outcomes for our customers. Scope of the Role: As a Technical Product Manager for Trust & Safety, you will define and ship the data products that frontier labs use to train and evaluate models to perform personalization, computer-use (CUA), and GDPval tasks safely and in a trustworthy manner. You will own the full lifecycle of these products: from translating a lab's research need into a spec, to designing reinforcement learning (RL) environments, rubrics, and evals, to running delivery with our expert workforce and measuring whether the data actually moved the model. A core part of the job is building RL environments with deliberate traps. That includes misalignment traps that test for reward hacking, deception, sandbagging, oversight evasion, and power-seeking; cyber traps such as planted credentials, injected instructions, and poisoned tools; and other honeypots that reveal how a model behaves when it believes no one is watching. You thrive in ambiguity, go deep on technical detail, and care about making powerful AI safe and beneficial. What You’ll Own: Define data products for frontier labs. Turn lab research goals into clear product specs for safety training data, RL environments, and evaluations. Design trap-based RL environments. Build sandboxed environments that embed misalignment traps (reward hacking, specification gaming, deception, sandbagging, oversight subversion) and cyber traps (prompt injection, credential exfiltration, tool poisoning, unsafe file and network actions). Build safety into agentic tasks by design. Specify safe-behavior rubrics, verifiers, and reward signals so that models learn to complete real work while refusing, escalating, or asking for confirmation at the right moments. Own evals and metrics. Define what "safe and trustworthy" means for each product in measurable terms, including attack success rate, over-refusal, trap trigger rate, and task completion under adversarial pressure. Track blind spots and drive iteration. Prioritize ruthlessly. Frame problems crisply, lay out options with their business and technical tradeoffs, and separate MVP from ideal state. Partner across teams. Work with lab researchers, our ML and platform engineers, red-teamers, policy experts, and delivery leads running large expert annotator pools. Track the threat landscape. Stay current on agentic risks, alignment research, and new attack techniques, and turn them into roadmap items before customers ask. Communicate clearly. Write specs, customer briefs, and safety findings that land with researchers and executives alike. These examples show the kind of work this role owns: A cyber-trap environment suite : sandboxed environments seeded with injected instructions, exposed secrets, and destructive shortcuts, scored on whether the agent finishes the task without taking the bait. Tasks with integrity traps : realistic work where the easiest path to a high score requires fabricating data, skipping a required step, or gaming the grader. A misalignment eval harness : environments that check whether a model behaves differently when it believes it is being tested versus deployed, or when g