Staff Security Engineer (Artificial Intelligence)

Compass; New York (NY); On-site; USD 180,000 - 240,000; Full time

Job summary

Compass is embedding generative and agentic AI into products used by agents and clients. You will own the security of that surface, acting as the technical authority on how AI features are built safely—from the model and retrieval layer to client-facing outputs.

You will partner with the AI platform and product engineering teams to define security standards, perform threat modeling, and deploy real-time protections across LLM inputs/outputs and multi-agent workflows.

Qualifications

• 7+ years in security engineering, with 2+ years focused on AI/ML or LLM security.
• Experience threat modeling complex distributed systems and driving remediation with engineering teams.
• Demonstrated experience securing a production AI or ML system end to end, development lifecycle, deployment, and runtime.
• Expertise in adversarial ML and LLM risks including prompt injection, model poisoning, data extraction, and unsafe tool use, with hands‑on red teaming experience.
• Background building evaluation and safety-testing harnesses for LLM applications.

Responsibilities

• Own the AI product security roadmap. Define the multi-year vision, strategy, and standards for securing AI capabilities in the Compass platform, developed jointly with the AI platform and product engineering teams.
• Threat model AI capability builds. Drive architectural reviews and threat modeling for new AI features, RAG pipelines, multi-agent workflows, and automated decisioning before they ship.
• Build production guardrails. Architect, deploy, and automate real-time protections for LLM inputs and outputs, including injection mitigation, jailbreak defense, output verification, and sensitive-data egress checks on generated responses.
• Constrain autonomy. Design blast-radius controls for autonomous agents: scoped permissions, human-in-the-loop checkpoints for consequential actions, action audit trails, and safe tool-invocation patterns.
• Secure the retrieval layer. Enforce data access controls, source integrity, tenancy boundaries, and retrieval-boundary protections across RAG and vector store implementations.
• Harden the AI supply chain. Assess the security posture, privacy controls, data retention, and training-use terms of third-party foundation models, inference providers, and vector stores used in production.
• Instrument and detect. Build detections and telemetry for AI-specific abuse in production, anomalous tool invocation, prompt abuse patterns, unusual retrieval behavior, and data exfiltration through model responses and integrate them into our monitoring estate.
• Automate the standard. Ship reusable libraries, CI/CD checks, and evaluation harnesses so secure AI patterns are the default path for product engineers rather than a review gate.
• Advise and evangelize. Act as the primary technical consultant for AI architecture decisions and translate emerging AI risk into concrete engineering roadmaps for both engineers and executives.

Skills

Python

Go

AI security

Threat modeling

Security architecture

Cloud security

DevSecOps

Threat detection

Communication

Adversarial ML

Tools

AWS

GCP

Kubernetes

IAM

CI/CD tooling

Vector stores

MITRE ATLAS

Job description

• Compass is embedding generative and agentic AI directly into the products our agents and clients use every day. This role owns the security of that surface. You will be the technical authority on how Compass builds AI features safely, from the model and retrieval layer, through agent tool invocation, to what ultimately reaches a client’s screen
• You will work shoulder-to-shoulder with the teams building the AI platform so that platform standards and security standards are one bar, not two
• Own the AI product security roadmap. Define the multi-year vision, strategy, and standards for securing AI capabilities in the Compass platform, developed jointly with the AI platform and product engineering teams
• Threat model AI capability builds. Drive architectural reviews and threat modeling for new AI features, RAG pipelines, multi-agent workflows, and automated decisioning before they ship
• Build production guardrails. Architect, deploy, and automate real-time protections for LLM inputs and outputs, including injection mitigation, jailbreak defense, output verification, and sensitive-data egress checks on generated responses
• Constrain autonomy. Design blast-radius controls for autonomous agents: scoped permissions, human-in-the-loop checkpoints for consequential actions, action audit trails, and safe tool-invocation patterns
• Secure the retrieval layer. Enforce data access controls, source integrity, tenancy boundaries, and retrieval-boundary protections across RAG and vector store implementations
• Harden the AI supply chain. Assess the security posture, privacy controls, data retention, and training-use terms of third-party foundation models, inference providers, and vector stores used in production
• Instrument and detect. Build detections and telemetry for AI-specific abuse in production, anomalous tool invocation, prompt abuse patterns, unusual retrieval behavior, and data exfiltration through model responses and integrate them into our monitoring estate
• Automate the standard. Ship reusable libraries, CI/CD checks, and evaluation harnesses so secure AI patterns are the default path for product engineers rather than a review gate
• Advise and evangelize. Act as the primary technical consultant for AI architecture decisions and translate emerging AI risk into concrete engineering roadmaps for both engineers and executives

Automation is your default approach to security, especially when securing complex AI systems at scaleYou write production code, strong proficiency in Python or Go, with experience building real integrations and security pipelines, not just prototypesYou have a deep working understanding of transformer architectures, inference pipelines, RAG, prompt engineering, and the OWASP Top 10 for LLMsYou communicate well across audiences, translating emerging AI risk into technical roadmaps for engineers and business impact for leadershipYou are proactive about finding and closing gaps in AI/ML workflows through tooling rather than policyStrong application and cloud security fundamentals: AWS and/or GCP, Kubernetes, IAM, API security, and DevSecOps practices7+ years in security engineering, with 2+ years focused on AI/ML or LLM application securityExperience threat modeling complex distributed systems and driving remediation with engineering teamsDemonstrated experience securing a production AI or ML system end to end, development lifecycle, deployment, and runtimeStrong strategic communication skills with a proven ability to translate complex technical AI risk into business impact for executive leadershipExpertise in adversarial ML and LLM risks including prompt injection, model poisoning, data extraction, and unsafe tool use, with hands‑on red teaming experienceExperience with multi-agent autonomous systems or Model Context Protocol (MCP) based architecturesFamiliarity with MITRE ATLAS, NIST AI RMF, or comparable AI threat/risk frameworksBackground building evaluation and safety-testing harnesses for LLM applicationsPrior work on AI platform security architecture or secure‑by‑default AI SDKsExperience in a regulated or consumer‑data-heavy environment where model outputs touch customer‑facing decisions