OpenAI is spending millions to commit felonies | Eli the Computer Guy
Summary
Isaac and Eli the Computer Guy discuss recent "rogue" AI agent incidents, arguing that these failures stem from gross negligence and a lack of basic security protocols rather than advanced alignment issues. They criticize Anthropic and OpenAI for treating existential security threats as marketing opportunities, noting that standard tech stack testing and zero-trust environments could prevent such breaches. The hosts also debunk the feasibility of IoT-based AI worms and warn against the risks of agentic commerce, emphasizing that probabilistic models are ill-suited for handling sensitive fin…
Topics discussed
Sponsors: Citizens Bank and Perk
Intro: Rogue AI agents and lab negligence
Anthropic disables live internet access for models
Critique of frontier labs' basic security failures
Standard tech stack testing vs. AI development
Financial cost of rogue agents and token burn
OpenAI's recent breaches and PR strategy
Marketing danger as a feature to investors
Implementing visibility and network controls
Sponsors: Citizens Bank, Perk, Samaritan's Purse
Security risks of GrokBot and credit card uploads
Meta Muse accessing iMessage files and permissions
Probabilistic models and executive AI incidents
Appropriate use cases: decision support vs. action
Alignment narrative vs. zero-trust system design
Sponsors: Citizens Bank, Perk, Samaritan's Purse
Automated grading, throttling, and supervisory agents
Feasibility of sharded LLMs on IoT devices
Outro and final sponsor reads
Listen ad-free on Castria