Ezra Tanzer is a director of product management at Snyk focused on securing AI coding agents as they gain the ability to modify software, invoke external tools, and act across organizational systems. His work treats agent security as three connected problems: controlling what agents generate, what they use, and what they do.
Earlier in his career, Tanzer worked in product management at Dispatch and subsequently joined Topcoat Data, whose technology Snyk acquired. At Snyk, he initially focused on helping security leaders understand application exposure, remediation, prevention, and coverage. His writing on application-security reporting for CISOs translated vulnerability data into decisions about investment and accountability.
As coding assistants became commonplace, Tanzer shifted toward preventing insecure software before it enters review pipelines. His work on developer trust and fix fatigue argues that late alerts, poor explanations, and unreliable fixes cause developers to disengage from security tools. Useful safeguards must identify problems early and offer remediation worth accepting.
- Asynchronous, hooks-based guardrails. Tanzer favors security hooks that scan files immediately after an agent changes them, record newly introduced vulnerabilities outside the model’s context window, and initiate remediation when the session ends. This approach avoids relying on instructions an agent might ignore while reducing latency, token consumption, and unnecessary interruptions. Snyk Studio applies that architecture to agentic development.
- Agent supply-chain security. MCP servers, skills, extensions, and dependencies create risks beyond vulnerable generated code: excessive privileges, prompt injection, malicious instructions, and persistent changes to agent behavior. Tanzer has also examined how AI coding tools revive neglected open-source packages, making provenance and maintenance essential security checks.
- Enforceable boundaries outside the model. Tanzer’s guidance on the Agent Baseline distinguishes the controls needed for individual coding agents, shared internal assistants, and production systems acting on customer data. He advocates scoped access, observable execution, and policies that can block or redirect dangerous actions without depending on an agent’s judgment.
At AI Engineer World’s Fair 2026, Tanzer outlined this framework in a joint session with Snyk engineer Dan Arpino. Arpino separately demonstrated his own experimental monitoring tool. Tanzer’s product challenge is making autonomous development governable without burying developers in false positives or approval requests.
