AI Agent Security Architect
Designs and secures AI agent architectures, runtime defenses, and sandboxing frameworks for an agentic software creation platform. Work focuses on Model Context Protocol, vector stores, and mitigating non-deterministic AI threats.
- Compensation
- $230–$350 / yr
- Location
- Foster City, CA
- Work setup
- Hybrid
- Seniority
- Not specified
- Employment
- Full time
- Experience
- 2+ years
- Category
- Security
- Posted
Skills
Must have
- 6+ years in security engineering or architecture
- 2+ years in AI/ML security or LLM application security
- Deep understanding of agentic architectures and protocols
- Hands-on expertise with agentic AI threat vectors
- Strong background in AI safety standards and risk frameworks
- Proficiency in Python, Go, or TypeScript/JavaScript
- Experience with security architecture documentation
- Exceptional communication of complex AI risks
- Experience contributing to Cybersecurity Risk Register
Nice to have
- Experience with runtime sandboxing and isolation technologies
Full description
Replit is the agentic software creation platform that enables anyone to build applications using natural language. With millions of users worldwide, Replit is democratizing software development by removing traditional barriers to application creation.
About the Role
We are looking for an AI Agent Security Architect to function as the primary technical authority for Replit’s autonomous and AI agent security blueprint. In this critical role, you will design, implement, and maintain the runtime defense systems, guardrail frameworks, and sandboxing architectures that govern AI agents executing code, invoking tools, and reasoning across our platform. You will be a key technical contributor—leading high-impact AI security initiatives and bridging the gap between non-deterministic AI behavior and rigorous cybersecurity controls for both engineering and executive leadership.
What You'll Do
- AI Agent Security Strategy & Technical Execution
- AI Agent Security Blueprint: Define the long-term vision and architectural patterns for securing autonomous agent workflows, Model Context Protocol (MCP) integrations, multi-turn reasoning loops, and multi-agent coordination.
- Runtime Guardrails & Policy Enforcement: Architect and deploy dynamic input/output guardrail systems, semantic firewalls, and real-time intent verification filters to prevent goal hijacking, system prompt leaks, and indirect prompt injections.
- Agent Execution & Tool Sandboxing: Partner with Infrastructure and AppSec teams to design secure, short-lived, micro-isolated environments (e.g., microVMs, WebAssembly, container sandboxes) where agents can dynamically execute code, run shell commands, and interact with host operating systems safely.
- Agentic Threat Modeling & Red Teaming: Conduct specialized threat modeling against non-deterministic systems. Lead automated and manual AI red-teaming initiatives to uncover vulnerabilities in RAG context pipelines, vector stores, and tool-calling interfaces.
- Identity, Delegation & Least Privilege: Architect fine-grained permission models and step-up Human-in-the-Loop (HITL) authorization patterns for agents acting on behalf of users—ensuring agents never exceed the delegated privileges of the end user or misuse API keys.
- Context & Memory Boundary Security: Design strict data isolation and poisoning prevention controls for vector databases, retrieval-augmented generation (RAG) pipelines, and long-term conversation memories across multi-tenant workloads.
- Risk Management & Cross-Functional Enablement
- Maintain the AI Security Source of Truth: Define, document, and maintain authoritative secure design patterns and SDKs for engineering teams building internal or customer-facing AI agent capabilities.