Agent Security Research Engineer – Taiwan

Agent Security Research Engineer – Taiwan

About the role: You'll own agent security content from idea to production. You'll research how AI agents (coding assistants, desktop agents, workflow automation tools, and MCP-connected tools) can misbehave or be compromised. Then you'll design the detection or policy, build it in our engines and pipelines, test it, ship it, document it, help customers adopt it, and tune it based on real-world results. You're accountable for each piece of content delivering security value in customers' environments.

Required Skills:

Hands-on knowledge of how agentic tools work: tool calling, hooks and lifecycle events, MCP servers and transports, skills and plugins, permission modes

Solid grasp of agent-specific attack classes: direct and indirect prompt injection, tool and description poisoning, confused-deputy and excessive-agency issues, secret leakage into context, and malicious or over-permissioned MCP servers and extensions.

Able to tell real risk from theoretical risk and explain the difference to a customer.

Strong SQL, ideally on analytical stores like ClickHouse, Databricks, or Snowflake.

Disciplined about testing: builds reproducible test cases (positive and negative) before shipping content.

Proficient scripting in Python or Go, enough to build test harnesses, parse event logs, and perform detections.

Working knowledge of SaaS and cloud identity (OAuth, PATs, API tokens, scopes) and of developer tooling (Git, GitHub/GitLab, CI/CD).

Able to explain an agent attack chain to engineers, product and customer-facing teams.

Nice to Have:

Published research, CVEs, talks, or blog posts on LLM or agent security.

Familiarity with threat detection.

Hands-on experience with endpoint security on macOS, Linux, or Windows

Experience with dbt, Dagster, or other data-pipeline tooling.

Background in a SaaS security, CASB/SSPM, or insider-threat product.

Has used Claude Code, Copilot, Cursor, or similar heavily in daily work and has opinions about their security models.