Risky Business #854 -- We're Jevpilled
Summary
This episode covers Gemini's minor security lapses, OpenAI's new alignment reporting framework, and a critical sandbox escape in Codex. The hosts also discuss a significant breach of OpenAI's Discourse forum, US Treasury Secretary Scott Bessent's stance on AI liability, and the rise of the "Jev" decision model. Additional topics include a military operation aborted due to a chatbot hallucination, new zero-days in Cisco and WordPress, and the takedown of the Evil Tokens phishing platform.
Topics discussed
Introduction and sponsor review
Gemini's 'crimes' and LLM behavior
OpenAI agent misalignment disclosures
Codex sandbox escape bug details
Hackers' out-of-scope actions on OpenAI
AI liability and frontier lab responsibility
US-China AI competition and offline inference
Military operation aborted due to AI hallucination
Kevin Collier's article on AI cyber threats
Risky Business startup and design partners
JAX: Type-safe AI classifier for security
FBI and Coast Guard board hacked oil tankers
Cisco and WordPress zero-day vulnerabilities
North American threat actor targeting China/Russia
UN report on North Korean IT worker scams
Google Threat Intelligence and C2 disruption
Takedowns of DDoS and crypto laundering services
Cloudflare API abuse and Shiny Hunters vs Clop
FBI data breach, Oracle exploit, and asset forfeiture
Sponsor interview: Thinkst Canary's Agent Provocateur
Listen ad-free on Castria