AI Agent Security Researcher

We are looking for an experienced AI Agent Security Researcher to investigate and evaluate unsafe behaviours in autonomous AI agents. The role involves building controlled testing environments, developing detection and evaluation frameworks, and measuring how AI agents behave when faced with opportunities to deceive, bypass safeguards, or misuse tools and permissions. The position is fully remote (based in Israel) and part of a team advancing AI safety and cybersecurity research.

Mission

Own the Rogue Agents area and co-own Benchmarking. You will study how autonomous agents deviate from intended behaviour, including specification gaming, deception, evasion, unsafe tool use and privilege escalation, and build the environments and detectors that surface these failures and measure how often they occur.

Key responsibilities

• Design sandboxed environments and scenarios that elicit and capture rogue-agent behaviours in a controlled, observable way.
• Build detectors, telemetry and graders that identify when an agent deceives, evades oversight or misuses its tools.
• Develop benchmarks that quantify the frequency and severity of unsafe autonomous behaviour across models.
• Investigate failure cases in depth and translate them into reproducible regression tests.
• Collaborate with the offensive-cyber and financial-crime tracks on cross-cutting agent evaluations.

Required Skills & Experience (Must Have)

• M.A. / M.Sc. in Computer Science from an accredited university (degree completed; thesis track preferred).
• Prior hands-on experience at a cyber security company - a minimum of three years in a research, security engineering, threat-intelligence or fraud / financial-crime role at a company whose core business is cyber security.
• Hands-on experience building or securing AI agents, LLM tool-use pipelines or multi-agent systems.
• Solid software-engineering ability in Python, including building harnesses, sandboxes and evaluation scaffolding.
• Working knowledge of application and systems security, sandboxing and privilege boundaries.
• An experimental, adversarial mindset for probing and documenting failure modes.

Advantageous

• Research or publications in AI safety, agent evaluation or red-teaming of LLM systems.
• Experience with reinforcement-learning environments or agent frameworks.
• Contributions to open-source evaluation or agent-tooling projects.

What We Offer

• A frontline role shaping how the industry measures and governs AI cyber capability.
• A senior, collaborative research team and direct access to frontier models and tooling.
• Support for responsible publication of your research

About us

With more than 2,000 customers and operations in 13 cities around the world, Innodata is the AI technology solutions provider-of-choice for 7 of the world’s biggest technology companies, as well as leading companies across financial services, insurance, law, and medicine. We’re at an exciting stage in the company and fortunate in these times to grow our team.

We’re an equal opportunity employer.

Only relevant applicants will be contacted