I investigate security boundaries in AI agents, MCP integrations, and tool-enabled applications. My background is in Web3 security, including smart contracts, cross-chain authorization, bridge flows, and token-accounting invariants.
My current focus is AI Agent Security and AppSec: understanding how applications work, where trust changes, and whether security checks hold during execution.
- Authentication, authorization, tool permissions, and data isolation.
- Untrusted content and prompt-injection boundaries.
- MCP/OAuth sessions, token lifecycle, and transport errors.
- Shared state, concurrency, cancellation, and resource cleanup.
My portfolio includes local investigations of Browser Use, FastMCP / MCP Python SDK, and a LangGraph-based application. The work covers navigation-policy enforcement, error propagation after token revocation, and SQLite resource lifecycle, including reproductions, upstream fix verification, and a local cleanup patch with regression tests.
I participated in ETHOnline 2026 with Voxel Bench. The project publicly credits me for SR-01, a recipe-authorship vulnerability identified through AI-assisted contract review and local tests. The team implemented an attester-based fix and retained the regression evidence.
Map the architecture → define boundary invariants → design controls and negative scenarios → test locally → inspect evidence → verify fixes.
I use AI-assisted tooling for source analysis, test generation, and automation. Conclusions rely on observed behavior, logs, and repeatable checks, with scope and remaining evidence gaps recorded in each report.
Open to remote AI security and application security roles, contract work, and research collaboration.