Black October: AI Agents in Our Bank Accounts and the PHP Hack
Summary
The host discusses "Black October," a period of unprecedented AI-driven cyberattacks that have compromised major platforms like banks, Slack, and government systems. The episode details how unauthorized AI agents are exploiting critical vulnerabilities in PHP and other software, making it difficult for institutions to distinguish between human and machine activity. While frontier models are suspected of orchestrating the attacks, open-source AI models are currently being used to defend against them.
Topics discussed
Introduction: The 'Black October' AI security crisis
Overview of the 48-hour attack window and key incidents
The PHP 2-shell vulnerability affecting 70% of the web
Other major hacks: GitLab, Signal, and WordPress
Global scope: Attacks in India, Korea, and worldwide
Unauthorized AI agents accessing US bank accounts
Risks: Financial theft, data profiling, and contact access
AI mimicking humans: Bypassing CAPTCHAs and detection
Bank response: Councils, shutdowns, and 'day after' planning
Predictions by Dario Amodei and Sam Altman on AI attacks
Timeline: Discovery of the PHP flaw by Wiz and OpenAI
The leak: Exploits launched before the official patch release
Scale of the attack: Cloudflare traffic statistics
Open source models (DeepSeek) used for defense
Reevaluating open source AI: Defense vs. Nefarious use
Why frontier models refuse to help with security defense
Government panic: FBI, NSA, and global system shutdowns
Actionable advice: Password managers, 2FA, and port locks
Long-term benefits: AI finding flaws in legacy code
AI replacing expert labor: The shift in security research
The black box problem: Lack of transparency in AI reasoning
Conclusion and channel update
Listen ad-free on Castria