Agentic AI Security Engineer

Howard Hughes Medical Institute•Ashburn, VA

1d•$149,085 - $186,356•Onsite

About The Position

The Howard Hughes Medical Institute’s Janelia Research Campus is a pioneering research center in Ashburn, Virginia, where scientists pursue fundamental questions in the life sciences. Our integrated teams of biologists, computational scientists, and tool-builders innovate research practices and technologies to solve biology’s deepest mysteries. HHMI launched Janelia in 2006, establishing an intellectually enriching environment for scientists to do creative, collaborative, hands-on work. We share our methods, results, and tools with the scientific community. AI@HHMI at Janelia integrates AI agents across the research process. These tools write and execute code, optimize experimental and model parameters, and drive instruments. Letting AI agents act on shared research infrastructure requires serious safety engineering. The relevant failure mode is usually not a malicious attacker but a capable, misdirected tool running with legitimate user permissions at machine speed: hallucinated commands, instructions hijacked by retrieved content, runaway loops, sandbox escapes. We are hiring the engineer who owns this problem. You will research, develop, deploy, operate, and analyze the safety stack for agentic AI at Janelia: Sandboxing. Hardened, reproducible execution environments for agent code (kernel- and VM-level isolation, network egress control) on our OpenShift platform and HPC cluster, validated by adversarial testing. Guardrails. Policy enforcement on agent inputs, outputs, and tool calls; judge models; human approval gates; independent safeguard services that sessions can be routed through, individually or chained. Observability. Attributable audit logging, full-trace capture, real-time monitoring, and post-incident analysis of agent sessions. You will join early enough to make the foundational architecture decisions, and you will operate what you design. We expect and support open-sourcing and publishing this work.

Requirements

Hands-on experience operating LLM agents and dealing with their failures, or deep sandbox/container-security expertise with clear ability to move into the agentic domain.
Linux and Kubernetes/OpenShift isolation depth, including a precise understanding of the difference between containerization and sandboxing.
Working knowledge of agent architectures: tool-calling protocols (e.g., MCP), approval gates, judge models, guardrail frameworks, and the limits of each.
Evidence over certifications: open-source contributions, publications or technical writing, disclosed findings, or production systems you can discuss in depth.
Minimum 5 years in security, systems/SRE, or ML systems engineering.
Ability to see and hear within normal parameters.
Ability to move about workspace.
Ability to move materials weighing up to several pounds (such as a laptop computer or tablet).

Nice To Haves

Experience with Anthropic and Google collaborations.
Experience with NVIDIA GPU clusters (B300/H200/H100 class).

Responsibilities

Research, develop, deploy, operate, and analyze the safety stack for agentic AI at Janelia.
Develop and maintain sandboxing solutions, including hardened, reproducible execution environments for agent code (kernel- and VM-level isolation, network egress control) on OpenShift and HPC clusters, validated by adversarial testing.
Implement guardrails for policy enforcement on agent inputs, outputs, and tool calls, including judge models, human approval gates, and independent safeguard services.
Establish and manage observability systems for agentic AI, including attributable audit logging, full-trace capture, real-time monitoring, and post-incident analysis of agent sessions.
Make foundational architecture decisions for the agentic AI safety stack.
Operate the designed safety systems.
Support open-sourcing and publishing of the work.