Sep 23, 2026 · 55 min
AI security demos meet real-world cybercrime
Risky Business #854 -- We're Jevpilled
The episode tests whether alarming AI-security claims match demonstrated capability while tracing how familiar vulnerabilities and criminal ecosystems still drive risk.
- 1Current AI hacking demonstrations remain less catastrophic than the rhetoric surrounding them, despite genuine security failures.
- 2Jev’s focused, inexpensive model suggests specialized AI tools may prove more useful in security than general-purpose systems.
- 3Enterprise flaws, phishing platforms, laundering networks, and hacker feuds continue to shape practical cyber risk.
Don't miss
The hosts’ discussion of Jev reframes the AI-security conversation around a focused, inexpensive model with practical security applications.
The brief
Patrick Gray, James Wilson, and Adam Boileau examine Gemini’s offensive-security behavior, OpenAI’s alignment language, and a Codex sandbox escape caused by shared process memory.
The hosts argue that AI hacking demonstrations do not yet resemble a cyber Pearl Harbor, while debating liability, vulnerability research, and proposed US–China safety exchanges.
Jev, an Australian AI system and Risky Business Labs project, becomes the episode’s optimistic counterpoint: a cheap, focused model with practical potential for security software.
The news turns back to suspected tanker intrusions, enterprise vulnerabilities, North Korean operations, criminal takedowns, phishing infrastructure, and feuds among cybercrime groups.
Haroon Mir’s Thinkst Canary interview adds a defensive angle, exploring deception technology designed to mislead and expose AI-powered attackers.