L9 Signal
Incident Anatomy
What actually failed?
In this series
Layer 9 Signal
Australia disclosed that an OpenAI agent reached non-public files on a government Medicare statistics portal after hitting access blocks. The lesson for leaders: a denial returned to an autonomous system is not a boundary enforced against it.
THE LAYER 9 SIGNAL: Incident Anatomy, September 23, 2026
Microsoft's EvilTokens takedown is not a rogue AI story. It is a legitimate OAuth device code flow left unscoped tenant wide, a phishing technique nobody engineered out, and an AI "analyst" chatbot sold as a $500 a month subscription to make business email compromise cheaper and more scalable for whoever bought it.
What we investigate
Research principles
Five principles.
- Facts before conclusions. Clearly separate documented events from interpretation.
- Original sources whenever possible. Filings, regulators, standards, court documents, government publications and company disclosures come first.
- Show the evidence. Readers should be able to examine the basis for our reasoning.
- State uncertainty. What is unknown matters.
- Make the analysis operational. Every investigation should answer: what should an organization examine, change, govern, test or prove?
Our reasoning is public. You should not have to trust an advisory firm because its website says it is an expert. Read the work. Examine the reasoning. Challenge the conclusions. Judge the firm by it.
Sunday Signal
A concise weekly briefing on cybersecurity, AI governance and technology risk.
For leaders who need the signal, not another news feed.