Research Engineer, AI Security

Description

The company does cybersecurity for a world where AI writes much of the code. Coding agents now ship code faster than teams can review or secure it, with 100+ pull requests per engineer per week, and attackers use the same agents to find and exploit weaknesses within days. Quarterly pentests and long alert queues cannot keep pace.

The company runs AI security agents that reason over a live graph of a customer's code, CI, cloud and telemetry, and continuously attack a faithful copy of that system. Each finding comes validated, with its blast radius and a remediation plan that fits the customer's workflow so a coding agent can apply the fix. Integration takes hours, and the first validated finding arrives within days.

The founders previously built a security company that Datadog acquired. The company is backed by Accel and Y Combinator and is early in its growth.

Responsibilities

• Your job is to make the agents' judgment measurably better: how they plan, reason over a customer's production graph, choose what to test, and decide whether a finding is real.
• You sit between research and engineering. One week could include designing an eval that catches a new type of false positive, changing how an agent plans a multi-step investigation, benchmarking models on a hard task, and hardening agents against prompt injection from the systems they inspect. You will work with the founders, engineers, and the security leads who run pentests with the agents, and your work goes to production.
• Improve how agents plan, use tools, and reason over a live graph of code, cloud and telemetry, so they reach real findings sooner with fewer dead ends.
• Build evals, benchmarks and graders that show whether an agent improved, and make them a standard step in every change.
• Compare models, prompts and agent designs on cost, latency and quality, then turn the results into production choices.
• Protect agents against prompt injection, data exfiltration and manipulation by the systems under test. Research the security of LLM applications that customers build.
• Keep long-running agents dependable across thousands of runs through verification, self-checks and sound stopping rules.
• Review agent runs from customer engagements, find where reasoning failed, and turn each failure into an eval and a fix.

Requirements

• You have built, evaluated or improved LLM agents or systems running in production.
• You run careful experiments, do not trust a single anecdote, and let the evidence decide.
• You write production-grade code in Python, TypeScript or Go, and build tools others depend on.
• You understand security well enough, or learn it quickly: how real systems get attacked and how findings get validated.
• You know LLM security risks such as prompt injection, jailbreaks and misuse of tools.
• You use AI coding agents such as Claude Code or Cursor every day and have clear views on what makes them dependable.
• You have a background in machine learning, AI safety or security, shown through publications, open-source work or talks.
• You have offensive security experience: pentesting, red teaming, bug bounties or CTFs.
• You have worked with reinforcement learning, fine-tuning or synthetic data generation for agents.
• You have worked with graph data models or with reasoning over structured data.
• You have worked at an early-stage startup before.

Conditions

Salary is competitive and depends on level and experience.

Based in San Francisco, working in person.

We are an equal opportunity employer and welcome applicants from every background.

Benefits

• Early-stage equity in what the company builds.
• Comprehensive medical, dental and vision coverage.
• The equipment you need to do your best work.