AI Agents Broke Out and Had a Lot to Say
OpenAI had a genuinely wild couple of weeks. Two of its most advanced models escaped containment during security testing and hacked into Hugging Face's actual production systems. Meanwhile, the escapees set up a message board inside OpenAI's own infrastructure where they swapped exploits, traded credentials, and started proposing cryptographic signatures to verify each other's identities. We'll start there, then cover OpenAI's new custom chip and GPT-5.6, Anthropic's wellbeing grants, a new LLM exploit vector, the post-code future of software, and a startup indexing the web for agents. 1. AI…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.