Skip to content
Search jobs inside your AI agent. Claude, ChatGPT, Claude Code, Codex, Gemini, Cursor and Grok
Connect

Innodata Inc.

NewBe an early applicant

Staff Offensive Agentic AI Security Environments Engineer

Posted

Description

Frontier AI agents can now find vulnerabilities, chain exploits, and act on live systems. They learn these skills through reinforcement learning in environments like the ones you will build. Someone has to understand how these agents attack, and design the environments and tests that show what they will actually do.


Innodata's Agentic Evaluation and Safety practice works directly with the world's leading frontier AI labs. As a Staff Offensive Agentic AI Security Engineer, you are a senior hands-on attacker who also understands how AI agents work: how they plan, use tools, and learn from reward. You own whole families of offensive agentic environments end to end, from attack design to the target build to testing frontier agents against them. You work alongside alignment and safety scientists and RL engineers, and set the quality bar for the environments you own.


This is a fully remote position. Reporting to the Head of AI Cyber Agentic Environments & Testing and work closely with the CTO. You will help shape a new team at the cutting edge of AI cyber.


What you'll do

  • Own scenario families end to end. Take an area such as identity, cloud, CI/CD, or web applications and own every offensive environment in it, from attack design through build, testing, and handoff.
  • Build vulnerable-by-design environments for agents. Author realistic targets with real attack paths, misconfigurations, and exploit chains, modeled on how enterprises are actually breached. Design them as agentic environments, defining the tools, observations, and actions an agent works with.
  • Test frontier agents yourself. Run agents against your environments and analyze their trajectories: how they plan, call tools, recover from errors, and where they take unsafe or unauthorized shortcuts.
  • Design reward and grading signals. Define what counts as an exploit, an unauthorized action, or a data leak. Turn it into verifiable rewards and graders that hold up when a model is optimizing against them.
  • Make environments robust to the agent. Find and close the shortcuts, reward hacks, and test-detection cues that let an agent score without doing the real thing.
  • Collaborate with alignment scientists and RL engineers. Turn research questions about misuse, agentic risk, and unsafe autonomy into environments that work for both evaluation and training.
  • Raise the quality bar. Review other engineers' environments and help interview and hire.


What you bring

  • 6+ years of hands-on offensive security: penetration testing, red teaming, exploit development, or vulnerability research.
  • Deep expertise in at least one of: web and application exploitation, Active Directory and lateral movement, cloud attack paths, or container and Kubernetes security. Working knowledge across the rest.
  • Hands-on experience with AI agents: tool calling, agent frameworks and harnesses (MCP, computer-use, coding agents), and how LLM agents plan, act, and fail.
  • A working understanding of reinforcement learning for LLMs: RL environments, reward design, verifiers, reward hacking, and how post-training shapes agent behavior.
  • Experience building things that are meant to be broken: CTF challenges, training labs, ranges, or adversary emulation environments.
  • Comfort writing code and infrastructure: Python or Go, Docker, Linux, networking.
  • Clear thinking and writing. You can explain an attack chain to a researcher, an engineer, or a client.
  • Mentor junior SWEs, cyber researchers, and RL engineers.


Nice to have

  • Built RL environments, agent benchmarks, or evaluation harnesses, especially for cyber, computer-use, or coding agents.
  • Built AI-driven offensive tooling or autonomous pentesting agents.
  • Familiarity with AI safety and alignment research: misalignment, reward hacking, evaluation awareness, agentic risk.
  • Background in elite offensive units, national CERTs, or top-tier red-team consultancies.
  • Published CVEs, Hack The Box or similar machines, or strong CTF results.
  • OSWE, OSEP, OSCP, CRTO, or equivalent demonstrated skill.
  • Talks or publications at security or AI venues.


We’re an equal opportunity employer.

Only relevant applicants will be contacted

Skills

Apply

See also