Intellectual Architecture

Archive of Insights.

cybersecurity / 2026

The Agents Built a Message Board: Inside OpenAI's Collective

At Black Hat, OpenAI walked through what actually happened in July. It did not start in July, and it was not one agent. Agents in the training fleet found each other, built a shared message board, and escalated together.

Read Analysis arrow_forward
psychology microsoft

Project Perception: Microsoft's Answer to Machine-Speed Attacks

Microsoft's Project Perception answers a year of AI-driven attacks with red, blue and green agents in a closed loop, and a multi-model system Microsoft says outscores the frontier at half the cost.

Read Post

The Call Is Coming From Inside the House: When Your Own Agent Becomes the Attacker

An AI agent escaped a cyber evaluation and compromised another company's production systems. The owner-side attribution failure matters as much as the exploit chain.

visibility 23 MIN READ

Outside the Model: Governing AI Agents on the Microsoft Stack

A practical control map for governing AI agents with Entra, Defender, Global Secure Access and Purview, including GA/preview boundaries and custom detection.

visibility 18 MIN READ

When Effort Became Free: AI, Mythos, and the New Economics of Hacking

One person used Claude Code while exploiting a flaw in Ajax's ticketing app. Lower marginal effort changes which targets are worth attacking and how defenders must respond.

visibility 11 MIN READ

Blueprint for a Modern Defense Stack: Why Attackers Have a Better Map Than You

Raviv Tamir's Experts Live Denmark keynote showed why attackers see connected paths while defenders still work through disconnected lists.

visibility 6 MIN READ

From Chat to Code to Claw: The Three Ages of AI Interaction

Chat, Code and Claw describe three overlapping patterns of human-AI interaction: asking for an answer, delegating a job and running a persistent agent.

visibility 7 MIN READ

The Odido Breach: Anatomy of a 6.5 Million-Record Data Leak and How to Defend Against It

A fluent Dutch caller impersonated IT, compromised a helpdesk account and reached Odido's Salesforce data. This is what the public evidence and defensive controls show.

visibility 11 MIN READ