Lead AI Security Architect: Platform Threat & Defense

EY; Dallas (TX); Remote; USD 140,000 - 190,000; Full time

Job summary

EY seeks an AI Security Engineer to own the security posture of its Agentic AI platform end to end, safeguarding from cloud to air-gapped environments. You will define threat models, enforce secure-by-default contracts, and lead red-team exercises that test guardrails and authority boundaries.

The role spans model provenance, supply-chain integrity, and policy-as-code for enterprise clients. The position requires deep cloud-native security expertise, experience with AI threat models, and the

Responsibilities

• Own the platform threat model covering agent autonomy, tool invocation, delegated authority, model and data supply chain, multi-tenancy, and every deployment target from cloud to air-gapped.
• Define the security engineering and control set for every platform layer: infrastructure and boot chain, Kubernetes and cluster fabric, identity and secrets, secure execution and sandboxing, gateway and egress, data and state, delivery pipeline, and telemetry.
• Set the secure-by-default contract so that platform capabilities arrive hardened, including agent templates, Helm charts, sandbox profiles, and network policy ship with correct controls.
• Own defense against agentic threat classes including prompt injection, jailbreak and instruction hijacking, excessive agency, confused-deputy and authority-escalation attacks, tool and function-call abuse, memory and context poisoning, and retrieval-augmented data exfiltration.
• Work with the architecture team to help define the agent authority model: delegated and on-behalf-of authority, scope and delegation-depth limits, consent boundaries, and the non-escalation invariant that an agent never exceeds the authority of its initiating principal at any hop.
• Own the sandboxing security standard for agent-generated code execution: isolation boundaries, filesystem and credential scope, egress restriction, resource containment, and the escape-test suite that proves the boundary holds.
• Secure the model and knowledge supply chain: model provenance and integrity, upstream registry governance, poisoning and backdoor risk, embedding and vector-store integrity.
• Secure agent-to-agent and tool protocols including MCP and A2A surfaces: discovery trust, tool registration and approval, schema validation, and authorization of inter-agent calls.
• Lead AI red teaming and adversarial testing: build the offensive capability and the recurring exercise cadence that tests guardrails, sandboxes, and authority boundaries before adversaries and auditors do.
• Own supply-chain integrity end to end: artifact signing and verification (Sigstore/Cosign, Notation), SBOM generation and attestation, provenance and SLSA-aligned build integrity, CVE management, dependency and license governance.
• Own admission and runtime policy: policy-as-code across Kyverno and OPA, signature-verification enforcement, Pod Security Standards, and the guardrails that make non-compliant workloads unschedulable rather than merely reported.
• Define Kubernetes and infrastructure hardening baselines: CIS-aligned cluster configuration, network default-deny and segmentation, node and boot-chain integrity, GPU and DPU isolation, and secrets-handling standards.
• Own tenant isolation assurance: the security definition of a tenant boundary across compute, network, storage, secrets, telemetry, and evidence, and the testing that proves cross-tenant leakage is not possible.
• Serve as the security authority in client

Job description

EY seeks an AI Security Engineer to own the security posture of its Agentic AI platform end to end, safeguarding from cloud to air-gapped environments. You will define threat models, enforce secure-by-default contracts, and lead red-team exercises that test guardrails and authority boundaries.

The role spans model provenance, supply-chain integrity, and policy-as-code for enterprise clients. The position requires deep cloud-native security expertise, experience with AI threat models, and the