Hackers On Planet Earth

It’s 10 PM. Do You Know What Your AI Agents Are Doing?

Alexander Rodriguez

AI coding agents like Claude Code now have shell access, file system access, and connections to external services through MCP servers – and most security teams have zero visibility into what they’re actually doing on developer machines. Alexander Rodriguez walks through building a three-layer open-source defense stack from scratch: an OpenTelemetry pipeline that captures every command, file access, MCP server connection, and permission decision; Meta’s LlamaFirewall wired into pre-execution hooks to block prompt injection and goal hijacking before actions run; and a lightweight EDR-style detection agent that watches AI agent behavior the way other tools watch process behavior by signature matching for credential file reads and exfil chains, behavioral baselining for anomalies, and real-time blocking. The talk also covers what’s still broken, such as no visibility into model reasoning, MCP servers that can change behavior after vetting, and prompt injection detection limited to pattern matching. Full stack on GitHub, demo included.

Saturday 1900 Gramercy Park Suite

Alexander Rodriguez is a senior security engineer focused on detection and response. He builds detection pipelines, incident response automation, and endpoint security tooling, and is currently developing open-source monitoring and prevention frameworks for AI coding agents using OpenTelemetry and Claude Code hooks. He has spent his career in defensive security across biotech, e-commerce, healthcare, and higher education.