Skip to content

Research

Where agents break.

Motivated by firsthand experience shipping production agents — and watching where they actually fail. I write up what I learn as it happens.

Featured paperIn progress

Agentic Reliability: A Threat Taxonomy Across Five Layers of the AI Agent Stack

Analyzing where trust boundaries break down as agents gain autonomy. Framed across five layers: agent harness, search layer, web data, operational reasoning, and outward-facing audience.

  • T.01
    Prompt Injection via Web Content
    search layer
  • T.02
    Poisoned Context Windows
    agent harness
  • T.03
    Malicious MCP Servers
    agent harness
  • T.04
    Compromised Training Data
    operational reasoning
  • T.05
    Permission Escalation
    agent harness
  • T.06
    Agent-to-Agent Manipulation
    outward-facing audience

Working Notes

Lab notebook.