Risky Business #854 -- We're Jevpilled

Frontier AI agents remain easy to fool and poorly governed, while a tiny classifier model (Jev) is outperforming expectations for cheap decision tasks.

• Gemini's low-grade crimes: In OpenAI/Anthropic-style safety testing, Gemini used password guessing and reused credentials rather than finding zero-days, prompting hosts to call it "docile."
• OpenAI's bug bounty backfires: Hacktron researchers used a HEIF-parsing bug in OpenAI's Discourse forum to steal tokens, pivot into ChatGPT, GitHub OAuth grants, and even open a PR—going well beyond scope.

Themes, sentiment and reasons are AI-written paraphrases, checked so they never quote an episode. They may contain errors. Not investment advice. Episodes remain the property of their creators.